GitLab

In GICv4, we will ensure that level2 vPE table memory is allocated
for the specified vpe_id on all v4 ITS, in its_alloc_vpe_table().
This still works well for the typical GICv4.1 implementation, where
the new vPE table is shared between the ITSs and the RDs.

To make it explicit, let us introduce allocate_vpe_l2_table() to
make sure that the L2 tables are allocated on all v4.1 RDs. We're
likely not need to allocate memory in it because the vPE table is
shared and (L2 table is) already allocated at ITS level, except
for the case where the ITS doesn't share anything (say SVPET == 0,
practically unlikely but architecturally allowed).

The implementation of allocate_vpe_l2_table() is mostly copied from
its_alloc_table_entry().
Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200206075711.1275-4-yuzenghui@huawei.com

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

The Size field of GICv4.1 VPROPBASER register indicates number of
pages minus one and together Page_Size and Size control the vPEID
width. Let's respect this requirement of the architecture.
Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20200206075711.1275-2-yuzenghui@huawei.com

It looks like an obvious mistake to use its_mapc_cmd descriptor when
building the INVALL command block. It so far worked by luck because
both its_mapc_cmd.col and its_invall_cmd.col sit at the same offset of
the ITS command descriptor, but we should not rely on it.

Fixes: cc2d3216

 ("irqchip: GICv3: ITS command queue")
Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20191202071021.1251-1-yuzenghui@huawei.com

Just like for INVALL, GICv4.1 has grown a VPE-aware INVLPI register.
Let's plumb it in and make use of the DirectLPI code in that case.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191224111055.11836-16-maz@kernel.org

Since GICv4.1 gives us a per-VPE doorbell, avoid programming anything
else on VMOVI/VMAPI/VMAPTI and on any other action that would have
otherwise resulted in a per-VLPI doorbell to be programmed.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191224111055.11836-15-maz@kernel.org

GICv4.1 redistributors have a VPE-aware INVALL register. Progress!
We can now emulate a guest-requested INVALL without emiting a
VINVALL command.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191224111055.11836-14-maz@kernel.org

When descheduling a VPE, special care must be taken to tell the GIC
about whether we want to receive a doorbell or not. This is a
major improvement on GICv4.0, where the doorbell had to be separately
enabled/disabled.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191224111055.11836-13-maz@kernel.org

Making a VPE resident on GICv4.1 is pretty simple, as it is just a
single write to the local redistributor. We just need extra information
about which groups to enable, which the KVM code will have to provide.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191224111055.11836-12-maz@kernel.org

masking/unmasking doorbells on GICv4.1 relies on a new INVDB command,
which broadcasts the invalidation to all RDs.

Implement the new command as well as the masking callbacks, and plug
the whole thing into the v4.1 VPE irqchip.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191224111055.11836-11-maz@kernel.org

Just like for GICv4.0, each VPE has its own doorbell interrupt, and
thus an irqchip that manages them. Since the doorbell management is
quite different on GICv4.1, let's introduce an almost empty irqchip
the will get populated over the next new patches.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191224111055.11836-10-maz@kernel.org

With GICv4.1, VMOVP is extended to allow a default doorbell to be
specified, as well as a validity bit for this doorbell. As an added
bonus, VMOVP isn't required anymore of moving a VPE between
redistributors that share the same affinity.

Let's add this support to the VMOVP builder, and make sure we don't
issue the command if we don't really need to.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191224111055.11836-9-maz@kernel.org

The infamous VPE proxy device isn't used with GICv4.1 because:
- we can invalidate any LPI from the DirectLPI MMIO interface
- the ITS and redistributors understand the life cycle of
  the doorbell, so we don't need to enable/disable it all
  the time

So let's escape early from the proxy related functions.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191224111055.11836-8-maz@kernel.org

The ITS VMAPP command gains some new fields with GICv4.1:
- a default doorbell, which allows a single doorbell to be used for
  all the VLPIs routed to a given VPE
- a pointer to the configuration table (instead of having it in a register
  that gets context switched)
- a flag indicating whether this is the first map or the last unmap for
  this particular VPE
- a flag indicating whether the pending table is known to be zeroed, or not

Plumb in the new fields in the VMAPP builder, and add the map/unmap
refcounting so that the ITS can do the right thing.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191224111055.11836-7-maz@kernel.org

GICv4.1 defines a new VPE table that is potentially shared between
both the ITSs and the redistributors, following complicated affinity
rules.

To make things more confusing, the programming of this table at
the redistributor level is reusing the GICv4.0 GICR_VPROPBASER register
for something completely different.

The code flow is somewhat complexified by the need to respect the
affinities required by the HW, meaning that tables can either be
inherited from a previously discovered ITS or redistributor.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191224111055.11836-6-maz@kernel.org

While GICv4.0 mandates 16 bit worth of VPEIDs, GICv4.1 allows smaller
implementations to be built. Add the required glue to dynamically
compute the limit.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191224111055.11836-3-maz@kernel.org

When updating an LPI configuration, get_vlpi_map() may be passed a
irq_data structure relative to an ITS domain (the normal case) or one
that is relative to the core GICv3 domain in the case of a GICv4
doorbell.

In the latter case, special care must be take not to dereference
the irq_chip data as an its_dev structure, as that isn't what is
stored there. Instead, check *first* whether the IRQ is forwarded
to a vcpu, and only then try to obtain the vlpi mapping.

Fixes: c1d4d5cd

 ("irqchip/gic-v3-its: Add its_vlpi_map helpers")
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reported-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20200122085609.658-1-yuzenghui@huawei.com

The VLPI map is currently a mutex, and that's a bad idea as
this lock can be taken in non-preemptible contexts. Convert
it to a raw spinlock, and turn the memory allocation of the
VLPI map to be atomic.
Reported-by: Heyi Guo <guoheyi@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20191108165805.3071-12-maz@kernel.org

Obtaining the mapping ivformation for a VLPI should always be
done with the vlpi_lock for this device held. Otherwise, we
expose ourselves to races against a concurrent unmap.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20191108165805.3071-11-maz@kernel.org

We have so far always injected/cleared VLPIs using either
INT+SYNC or CLEAR+SYNC sequences, but that's pretty wrong
for two reasons:

- SYNC only synchronises physical LPIs
- The collection ID that for the associated LPI doesn't match
  the redistributor the vPE is associated with

Instead, send an {INT,CLEAR}+VSYNC for forwarded LPIs, ensuring
that the ITS synchronises against the virtual pending table.
Reported-by: Zenghui Yu <yuzenghui@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20191108165805.3071-10-maz@kernel.org

We have so far alwways invalidated VLPIs usinc an INV+SYNC
sequence, but that's pretty wrong for two reasons:

- SYNC only synchronises physical LPIs
- The collection ID that for the associated LPI doesn't match
  the redistributor the vPE is associated with

Instead, send an INV+VSYNC for forwarded LPIs, ensuring that
the ITS can properly synchronise the invalidation of VLPIs.

Fixes: 015ec038

 ("irqchip/gic-v3-its: Add VLPI configuration handling")
Reported-by: Zenghui Yu <yuzenghui@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20191108165805.3071-9-maz@kernel.org

Obtaining the mapping information for a VLPI is something quite common,
and the GICv4.1 code is going to make even more use of it. Expose it as
a separate set of helpers.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191027144234.8395-8-maz@kernel.org
Link: https://lore.kernel.org/r/20191108165805.3071-8-maz@kernel.org

Now that we have a copy of TYPER in the ITS structure, rely on this
to provide the same service as its->device_ids, which gets axed.
Errata workarounds are now updating the cached fields instead of
requiring a separate field in the ITS structure.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191027144234.8395-7-maz@kernel.org
Link: https://lore.kernel.org/r/20191108165805.3071-7-maz@kernel.org

Now that we have a copy of TYPER in the ITS structure, rely on this
to provide the same service as its->ite_size, which gets axed.
Errata workarounds are now updating the cached fields instead of
requiring a separate field in the ITS structure.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191027144234.8395-6-maz@kernel.org
Link: https://lore.kernel.org/r/20191108165805.3071-6-maz@kernel.org

Instead of caching the GICv4 compatibility in a discrete way, cache the
TYPER register instead, which can then be used to implement the same
functionnality. This will get used more extensively in subsequent patches.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191027144234.8395-5-maz@kernel.org
Link: https://lore.kernel.org/r/20191108165805.3071-5-maz@kernel.org

We currently don't make much use of the DirectLPI feature, and it would
be beneficial to do this more, if only because it becomes a mandatory
feature for GICv4.1.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191027144234.8395-4-maz@kernel.org
Link: https://lore.kernel.org/r/20191108165805.3071-4-maz@kernel.org

Waiting for a redistributor to have performed an operation is a
common thing to do, and the idiom is already spread around.
As we're going to make even more use of this, let's have a primitive
that does just that.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191027144234.8395-3-maz@kernel.org
Link: https://lore.kernel.org/r/20191108165805.3071-3-maz@kernel.org

We allocate the collection mapping on device creation, but somehow
free it on the irqdomain free path, which is pretty inconsistent
and has led to bugs in the past.

Move it to the point where we teardown the device, making the
alloc/free symetric.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Zenghui Yu <yuzenghui@huawei.com>
Link: https://lore.kernel.org/r/20191108165805.3071-2-maz@kernel.org

The its_cmd_block struct can either have u64 or __le64
data in it, so make a anonymous union to remove the
sparse warnings when converting to/from these.
Signed-off-by: Ben Dooks <ben.dooks@codethink.co.uk>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20191017112955.15853-1-ben.dooks@codethink.co.uk

On a system without Single VMOVP support (say GITS_TYPER.VMOVP == 0),
we will map vPEs only on ITSs that will actually control interrupts
for the given VM.  And when moving a vPE, the VMOVP command will be
issued only for those ITSs.

But when issuing VMOVPs we seemed fail to present the exact ITSList
to ITSs who are actually included in the synchronization operation.
The its_list_map we're currently using includes all ITSs in the system,
even though some of them don't have the corresponding vPE mapping at all.

Introduce get_its_list() to get the per-VM its_list_map, to indicate
which ITSs have vPE mappings for the given VM, and use this map as
the expected ITSList when building VMOVP. This is hopefully a performance
gain not to do some synchronization with those unsuspecting ITSs.
And initialize the whole command descriptor to zero at beginning, since
the seq_num and its_list should be RES0 when GITS_TYPER.VMOVP == 1.
Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/1571802386-2680-1-git-send-email-yuzenghui@huawei.com

When allocating a range of LPIs for a Multi-MSI capable device,
this allocation extended to the closest power of 2.

But on the release path, the interrupts are released one by
one. This results in not releasing the "extra" range, leaking
the its_device. Trying to reprobe the device will then fail.

Fix it by releasing the LPIs the same way we allocate them.

Fixes: 8208d170

 ("irqchip/gic-v3-its: Align PCI Multi-MSI allocation on their size")
Reported-by: Jiaxing Luo <luojiaxing@huawei.com>
Tested-by: John Garry <john.garry@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/f5e948aa-e32f-3f74-ae30-31fee06c2a74@huawei.com

We try to find a free LPI region in device's lpi_map and allocate them
(set them to 1) when we want to allocate LPIs for this device. This is
what bitmap_find_free_region() has done for us. The following set_bit
is redundant and a bit confusing (since we only set_bit against the first
allocated LPI idx). Remove it, and make the set_bit explicit by comment.
Signed-off-by: Zenghui Yu <yuzenghui@huawei.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>

Do not expose the ITS' VA (it appears in debugfs). Instead, record
the PA, which at least can be used to precisely identify the associated
irqchip and domain.
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Marc Zyngier <maz@kernel.org>

In its_vpe_init, when its_alloc_vpe_table fails, we should free
vpt_page allocated just before, instead of vpe->vpt_page.
Let's fix it.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Jason Cooper <jason@lakedaemon.net>
Cc: Marc Zyngier <marc.zyngier@arm.com>
Signed-off-by: Nianyao Tang <tangnianyao@huawei.com>
Signed-off-by: Shaokun Zhang <zhangshaokun@hisilicon.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>

Arguments are supposed to be ordered high then low.
Signed-off-by: Joe Perches <joe@perches.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Marc Zyngier <marc.zyngier@arm.com>
Link: https://lkml.kernel.org/r/ab5deb4fc3cd604cb620054770b7d00016d736bc.1562734889.git.joe@perches.com

Based on 1 normalized pattern(s):

  this program is free software you can redistribute it and or modify
  it under the terms of the gnu general public license version 2 as
  published by the free software foundation this program is
  distributed in the hope that it will be useful but without any
  warranty without even the implied warranty of merchantability or
  fitness for a particular purpose see the gnu general public license
  for more details you should have received a copy of the gnu general
  public license along with this program if not see http www gnu org
  licenses

extracted by the scancode license scanner the SPDX license identifier

  GPL-2.0-only

has been chosen to replace the boilerplate/reference in 503 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Alexios Zavras <alexios.zavras@intel.com>
Reviewed-by: Allison Randal <allison@lohutok.net>
Reviewed-by: Enrico Weigelt <info@metux.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190602204653.811534538@linutronix.de

Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

When we run several VMs with PCI passthrough and GICv4 enabled, not
pinning vCPUs, we will occasionally see below warnings in dmesg:

ITS queue timeout (65440 65504 480)
ITS cmd its_build_vmovp_cmd failed

The reason for the above issue is that in BUILD_SINGLE_CMD_FUNC:
1. Post the write command.
2. Release the lock.
3. Start to read GITS_CREADR to get the reader pointer.
4. Compare the reader pointer to the target pointer.
5. If reader pointer does not reach the target, sleep 1us and continue
to try.

If we have several processors running the above concurrently, other
CPUs will post write commands while the 1st CPU is waiting the
completion. So we may have below issue:

phase 1:
---rd_idx-----from_idx-----to_idx--0---------

wait 1us:

phase 2:
--------------from_idx-----to_idx--0-rd_idx--

That is the rd_idx may fly ahead of to_idx, and if in case to_idx is
near the wrap point, rd_idx will wrap around. So the below condition
will not be met even after 1s:

if (from_idx < to_idx && rd_idx >= to_idx)

There is another theoretical issue. For a slow and busy ITS, the
initial rd_idx may fall behind from_idx a lot, just as below:

---rd_idx---0--from_idx-----to_idx-----------

This will cause the wait function exit too early.

Actually, it does not make much sense to use from_idx to judge if
to_idx is wrapped, but we need a initial rd_idx when lock is still
acquired, and it can be used to judge whether to_idx is wrapped and
the current rd_idx is wrapped.

We switch to a method of calculating the delta of two adjacent reads
and accumulating it to get the sum, so that we can get the real rd_idx
from the wrapped value even when the queue is almost full.

Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Jason Cooper <jason@lakedaemon.net>
Signed-off-by: Heyi Guo <guoheyi@huawei.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>

its_irq_compose_msi_msg() may be called from non-preemptible context.
However, on RT, iommu_dma_map_msi_msg requires to be called from a
preemptible context.

A recent change split iommu_dma_map_msi_msg() in two new functions:
one that should be called in preemptible context, the other does
not have any requirement.

The GICv3 ITS driver is reworked to avoid executing preemptible code in
non-preemptible context. This can be achieved by preparing the MSI
mapping when allocating the MSI interrupt.
Signed-off-by: Julien Grall <julien.grall@arm.com>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>

Using list_add + list_sort to insert an element and keeping the list
sorted is a somewhat blunt instrument; one can find the right place to
insert in fewer lines of code than the cmp callback uses. Moreover,
walking the entire list afterwards to merge adjacent ranges is
overkill, since we know that only the just-inserted element may be
merged with its neighbours.
Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>

There's no reason to ask kmalloc() to zero the allocation, since all
the fields get initialized immediately afterwards. Except that there's
also not any reason to initialize the ->entry member, since the
element gets added to the lpi_range_list immediately.
Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>