In order to maximize the use of MSI, a PCIe controller will share
all MSI controllers. The patch changes "msi-parent" to refer to all
MSI controller dts nodes.
Signed-off-by: Minghuan Lian <Minghuan.Lian@nxp.com>
Signed-off-by: Hou Zhiqiang <Zhiqiang.Hou@nxp.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
In order to maximize the use of MSI, a PCIe controller will share
all MSI controllers. The patch changes msi-parent to refer to all
MSI controller dts nodes.
Signed-off-by: Minghuan Lian <Minghuan.Lian@nxp.com>
Signed-off-by: Hou Zhiqiang <Zhiqiang.Hou@nxp.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
"1" should be replaced by "l". This is a typo.
The patch is to fix it.
Signed-off-by: Minghuan Lian <Minghuan.Lian@nxp.com>
Signed-off-by: Hou Zhiqiang <Zhiqiang.Hou@nxp.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
"1" should be replaced by "l". This is a typo.
The patch is to fix it.
Signed-off-by: Minghuan Lian <Minghuan.Lian@nxp.com>
Signed-off-by: Hou Zhiqiang <Zhiqiang.Hou@nxp.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
The patch is to fix typo of the Layerscape SCFG MSI dts compatible
strings. "1" is replaced by "l".
Acked-by: Rob Herring <robh@kernel.org>
Signed-off-by: Minghuan Lian <Minghuan.Lian@nxp.com>
Signed-off-by: Hou Zhiqiang <Zhiqiang.Hou@nxp.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
Initialization of irq_fwd_mask was done using __raw_writel() which
happens to work for all cases except when using ARM BE8 which requires
writel() (with the proper swapping). Move the initialization of the
irq_fwd_mask till later when we have correctly defined our I/O
accessors.
Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
Make these const as they are only used during a copy operation. Done
using Coccinelle.
@match disable optional_qualifier@
identifier s;
@@
static struct mmp_intc_conf s = {...};
@ref@
position p;
identifier match.s;
@@
s@p
@good1@
position ref.p;
identifier match.s,f,c;
expression e;
@@
(
e = s@p
|
e = s@p.f
|
c(...,s@p.f,...)
|
c(...,s@p,...)
)
@bad depends on !good1@
position ref.p;
identifier match.s;
@@
s@p
@depends on forall !bad disable optional_qualifier@
identifier match.s;
@@
static
+ const
struct mmp_intc_conf s;
Signed-off-by: Bhumika Goyal <bhumirks@gmail.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
Make this const as it is only used in a copy operation.
Done using Coccinelle.
Signed-off-by: Bhumika Goyal <bhumirks@gmail.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
As KVM needs to know about the availability of GICv4 to enable
direct injection of interrupts, let's advertise the feature in
the gic_kvm_info structure.
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
Do a braindump of the way things are supposed to work.
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
Add the required interfaces to map, unmap and update a VLPI.
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
Add the required interfaces to schedule a VPE and perform a
VINVALL command.
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
When creating a VM, it is very convenient to have an irq domain
containing all the doorbell interrupts associated with that VM
(each interrupt representing a VPE).
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
A long time ago, GITS_CTLR[1] used to be called GITC_CTLR.EnableVLPI.
It has been subsequently deprecated and is now an "Implementation
Defined" bit that may ot may not be set for GICv4. Brilliant.
And the current crop of the FastModel requires that bit for VLPIs
to be enabled. Oh well... Let's set it and find out what breaks.
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
While the doorbell interrupts are usually driven by the HW itself,
having a way to trigger them independently has proved to be a
really useful debug feature. As it is actually very little code,
let's add it to the VPE irqchip operations.
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
After moving a VPE from a redistributor to another, we're still left
with a potential pending doorbell interrupt on the old redistributor.
That interrupt should be moved to the new one to be either cleared
or take, depending on what the hypervisor wishes to do.
So let's move it right after having execited VMOVP. This doesn't
add much cost in the !DirectLPI case (we trade a DISCARD for a MOVI),
and the cost of the DIRECTLPI case should be minimal (two extra MMIO
accesses).
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
When we don't have the DirectLPI feature, we must work around the
architecture shortcomings to be able to perform the required
maintenance (interrupt masking, clearing and injection).
For this, we create a fake device whose sole purpose is to
provide a way to issue commands as if we were dealing with LPIs
coming from that device (while they actually originate from
the ITS). This fake device doesn't have LPIs allocated to it,
but instead uses the VPE LPIs.
Of course, this could be a real bottleneck, and a naive
implementation would require 6 commands to issue an invalidation.
Instead, let's allocate at least one event per physical CPU
(rounded up to the next power of 2), and opportunistically
map the VPE doorbell to an event. This doorbell will be mapped
until we roll over and need to reallocate this slot.
This ensures that most of the time, we only need 2 commands
to issue an INV, INT or CLEAR, making the performance a lot
better, given that we always issue a CLEAR on entry, and
an INV on each side of a trapped WFI.
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
The normal course of action when allocating the ITS' view of a
device is to allocate the corresponding LPIs. But we're about
to introduce devices that borrow their interrupts from
some other entities.
So let's make the allocation optional.
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
When masking/unmasking a doorbell interrupt, it is necessary
to issue an invalidation to the corresponding redistributor.
We use the DirectLPI feature by writting directly to the corresponding
redistributor.
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
When we're about to run a vcpu, it is crucial that the redistributor
associated with the physical CPU is being told about the new residency.
This is abstracted by hijacking the irq_set_affinity method for the
doorbell interrupt associated with the VPE. It is expected that the
hypervisor will call this method before scheduling the VPE.
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
When a guest issues a INVALL command targetting a collection, it must
be translated into a VINVALL for the VPE that has this collection.
This patch implements a hook that offers this functionallity to the
hypervisor.
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
When a VPE is scheduled to run, the corresponding redistributor must
be told so, by setting VPROPBASER to the VM's property table, and
VPENDBASER to the vcpu's pending table.
When scheduled out, we preserve the IDAI and PendingLast bits. The
latter is specially important, as it tells the hypervisor that
there are pending interrupts for this vcpu.
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
V{PEND,PROP}BASER being 64bit registers, they need some ad-hoc
accessors on 32bit, specially given that VPENDBASER contains
a Valid bit, making the access a bit convoluted.
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
On activation, a VPE is mapped using the VMAPP command, followed
by a VINVALL for a good measure. On deactivation, the VPE is
simply unmapped.
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
When creating a VM, the low level GICv4 code is responsible for:
- allocating each VPE a unique VPEID
- allocating a doorbell interrupt for each VPE
- allocating the pending tables for each VPE
- allocating the property table for the VM
This of course has to be reversed when the VM is brought down.
All of this is wired into the irq domain alloc/free methods.
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
Add the basic GICv4 VPE (vcpu in GICv4 parlance) infrastructure
(irqchip, irq domain) that is going to be populated in the following
patches.
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
When a VLPI is reconfigured (enabled, disabled, change in priority),
the full configuration byte must be written, and the caches invalidated.
Also, when using the irq_mask/irq_unmask methods, it is necessary
to disable the doorbell for that particular interrupt (by mapping it
to 1023) on top of clearing the Enable bit.
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
In order to let a VLPI being injected into a guest, the VLPI must
be mapped using the VMAPTI command. When moved to a different vcpu,
it must be moved with the VMOVI command.
These commands are issued via the irq_set_vcpu_affinity method,
making sure we unmap the corresponding host LPI first.
The reverse is also done when the VLPI is unmapped from the guest.
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
Add the skeleton irq_set_vcpu_affinity method that will be used
to configure VLPIs.
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
Add the new GICv4 ITS command definitions, most of them, being
defined in terms of their physical counterparts.
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
Add a bunch of GICv4-specific data structures that will get used in
subsequent patches.
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Marc Zyngier <marc.zyngier@arm.com>
Pull clockevent updates from Daniel Lezcano:
- Add the new imx-tpm driver (Dong Aisheng)
- Remove DT deprecated binding for Renesas (Magnus Damm)
- Remove error message on memory allocation (Markus Elfring)
- Convert clocksource drivers to use %pOF
Now that we have a custom printf format specifier, convert users of
full_name to use %pOF instead. This is preparation to remove storing
of the full path string for each node.
Signed-off-by: Rob Herring <robh@kernel.org>
Cc: Daniel Lezcano <daniel.lezcano@linaro.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Marc Gonzalez <marc_gonzalez@sigmadesigns.com>
Cc: Maxime Coquelin <mcoquelin.stm32@gmail.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Cc: linux-arm-kernel@lists.infradead.org
Acked-by: Marc Gonzalez <marc_gonzalez@sigmadesigns.com>
Acked-by: Alexandre TORGUE <alexandre.torgue@st.com>
Signed-off-by: Daniel Lezcano <daniel.lezcano@linaro.org>
Commit 87e81786b1 ("x86/idt: Move early IDT setup out of 32-bit asm")
switched early_ignore_irq to use ENTRY. ENTRY aligns the code, so there
is no need for one more ALIGN right before the function.
And add one \n after the function to separate it from the data.
Signed-off-by: Jiri Slaby <jslaby@suse.cz>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Josh Poimboeuf <jpoimboe@redhat.com>
Link: http://lkml.kernel.org/r/20170831121653.28917-1-jslaby@suse.cz
Add CONFIG_INFINIBAND_EXP_USER_ACCESS that enables the ioctl
interface. This interface is experimental and is subject to change.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
In order to use the parsing tree, we need to assign the root
to all drivers. Currently, we just assign the default parsing
tree via ib_uverbs_add_one. The driver could override this by
assigning a parsing tree prior to registering the device.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Adding CQ ioctl actions:
1. create_cq
2. destroy_cq
This requires adding the following:
1. A specification describing the method
a. Handler
b. Attributes specification
Each attribute is one of the following:
a. PTR_IN - input data
Note: This could be encoded inlined for
data < 64bit
b. PTR_OUT - response data
c. IDR - idr based object
d. FD - fd based object
Blobs attributes (clauses a and b) contain their type,
while objects specifications (clauses c and d)
contains the expected object type (for example, the
given id should be UVERBS_TYPE_PD) and the required
access (READ, WRITE, NEW or DESTROY). If a NEW is
required, the new object's id will be assigned to this
attribute. All attributes could get UA_FLAGS
attribute. Currently we support stating that an
attribute is mandatory or that the specification size
corresponds to a lower bound (and that this attribute
could be extended).
We currently add both default attributes and the two
generic UHW_IN and UHW_OUT driver specific attributes.
2. Handler
A handler gets a uverbs_attr_bundle. The handler developer uses
uverbs_attr_get to fetch an attribute of a given id.
Each of these attribute groups correspond to the specification
group defined in the action (clauses 1.b and 1.c respectively).
The indices of these arrays corresponds to the attribute ids
declared in the specifications (clause 2).
The handler is quite simple. It assumes the infrastructure fetched
all objects and locked, created or destroyed them as required by
the specification. Pointer (or blob) attributes were validated to
match their required sizes. After the handler finished, the
infrastructure commits or rollbacks the objects.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
In this phase, we don't want to change all the drivers to use
flexible driver's specific attributes. Therefore, we add two default
attributes: UHW_IN and UHW_OUT. These attributes are optional in some
methods and they encode the driver specific command data. We add
a function that extract this data and creates the legacy udata over
it.
Driver's data should start from UVERBS_UDATA_DRIVER_DATA_FLAG. This
turns on the first bit of the namespace, indicating this attribute
belongs to the driver's namespace.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Add a new ib_user_ioctl_verbs.h which exports all required ABI
enums and structs to the user-space.
Export the default types to user-space through this file.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
When some objects are destroyed, we need to extract their status at
destruction. After object's destruction, this status
(e.g. events_reported) relies in the uobject. In order to have the
latest and correct status, the underlying object should be destroyed,
but we should keep the uobject alive and read this information off the
uobject. We introduce a rdma_explicit_destroy function. This function
destroys the class type object (for example, the IDR class type which
destroys the underlying object as well) and then convert the uobject
to be of a null class type. This uobject will then be destroyed as any
other uobject once uverbs_finalize_object[s] is called.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This patch adds macros for declaring objects, methods and
attributes. These definitions are later used by downstream patches
to declare some of the default types.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Different drivers support different features and even subset of the
common uverbs implementation. Currently, this is handled as bitmask
in every driver that represents which kind of methods it supports, but
doesn't go down to attributes granularity. Moreover, drivers might
want to add their specific types, methods and attributes to let
their user-space counter-parts be exposed to some more efficient
abstractions. It means that existence of different features is
validated syntactically via the parsing infrastructure rather than
using a complex in-handler logic.
In order to do that, we allow defining features and abstractions
as parsing trees. These per-feature parsing tree could be merged
to an efficient (perfect-hash based) parsing tree, which is later
used by the parsing infrastructure.
To sum it up, this makes a parse tree unique for a device and
represents only the features this particular device supports.
This is done by having a root specification tree per feature.
Before a device registers itself as an IB device, it merges
all these trees into one parsing tree. This parsing tree
is used to parse all user-space commands.
A future user-space application could read this parse tree. This
tree represents which objects, methods and attributes are
supported by this device.
This is based on the idea of
Jason Gunthorpe <jgunthorpe@obsidianresearch.com>
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This adds the DEVICE object. This object supports creating the context
that all objects are created from. Moreover, it supports executing
methods which are related to the device itself, such as QUERY_DEVICE.
This is a singleton object (per file instance).
All standard objects are put in the root structure. This root will later
on be used in drivers as the source for their whole parsing tree.
Later on, when new features are added, these drivers could mix this root
with other customized objects.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Switch all uverbs_type_attrs_xxxx with DECLARE_UVERBS_OBJECT
macros. This will be later used in order to embed the object
specific methods in the objects as well.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
In this ioctl interface, processing the command starts from
properties of the command and fetching the appropriate user objects
before calling the handler.
Parsing and validation is done according to a specifier declared by
the driver's code. In the driver, all supported objects are declared.
These objects are separated to different object namepsaces. Dividing
objects to namespaces is done at initialization by using the higher
bits of the object ids. This initialization can mix objects declared
in different places to one parsing tree using in this ioctl interface.
For each object we list all supported methods. Similarly to objects,
methods are separated to method namespaces too. Namespacing is done
similarly to the objects case. This could be used in order to add
methods to an existing object.
Each method has a specific handler, which could be either a default
handler or a driver specific handler.
Along with the handler, a bunch of attributes are specified as well.
Similarly to objects and method, attributes are namespaced and hashed
by their ids at initialization too. All supported attributes are
subject to automatic fetching and validation. These attributes include
the command, response and the method's related objects' ids.
When these entities (objects, methods and attributes) are used, the
high bits of the entities ids are used in order to calculate the hash
bucket index. Then, these high bits are masked out in order to have a
zero based index. Since we use these high bits for both bucketing and
namespacing, we get a compact representation and O(1) array access.
This is mandatory for efficient dispatching.
Each attribute has a type (PTR_IN, PTR_OUT, IDR and FD) and a length.
Attributes could be validated through some attributes, like:
(*) Minimum size / Exact size
(*) Fops for FD
(*) Object type for IDR
If an IDR/fd attribute is specified, the kernel also states the object
type and the required access (NEW, WRITE, READ or DESTROY).
All uobject/fd management is done automatically by the infrastructure,
meaning - the infrastructure will fail concurrent commands that at
least one of them requires concurrent access (WRITE/DESTROY),
synchronize actions with device removals (dissociate context events)
and take care of reference counting (increase/decrease) for concurrent
actions invocation. The reference counts on the actual kernel objects
shall be handled by the handlers.
objects
+--------+
| |
| | methods +--------+
| | ns method method_spec +-----+ |len |
+--------+ +------+[d]+-------+ +----------------+[d]+------------+ |attr1+-> |type |
| object +> |method+-> | spec +-> + attr_buckets +-> |default_chain+--> +-----+ |idr_type|
+--------+ +------+ |handler| | | +------------+ |attr2| |access |
| | | | +-------+ +----------------+ |driver chain| +-----+ +--------+
| | | | +------------+
| | +------+
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
+--------+
[d] = Hash ids to groups using the high order bits
The right types table is also chosen by using the high bits from
the ids. Currently we have either default or driver specific groups.
Once validation and object fetching (or creation) completed, we call
the handler:
int (*handler)(struct ib_device *ib_dev, struct ib_uverbs_file *ufile,
struct uverbs_attr_bundle *ctx);
ctx bundles attributes of different namespaces. Each element there
is an array of attributes which corresponds to one namespaces of
attributes. For example, in the usually used case:
ctx core
+----------------------------+ +------------+
| core: +---> | valid |
+----------------------------+ | cmd_attr |
| driver: | +------------+
|----------------------------+--+ | valid |
| | cmd_attr |
| +------------+
| | valid |
| | obj_attr |
| +------------+
|
| drivers
| +------------+
+> | valid |
| cmd_attr |
+------------+
| valid |
| cmd_attr |
+------------+
| valid |
| obj_attr |
+------------+
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
We should report the network header type in the work completion so that
the kernel can infer the right RoCE type headers.
Reviewed-by: Bryan Tan <bryantan@vmware.com>
Signed-off-by: Aditya Sarwade <asarwade@vmware.com>
Signed-off-by: Adit Ranadive <aditr@vmware.com>
Reviewed-by: Yuval Shaia <yuval.shaia@oracle.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
For RoCE, ib_init_ah_from_wc() can follow the path
ib_init_ah_from_wc() ->
rdma_addr_find_l2_eth_by_grh() ->
rdma_resolve_ip()
and rdma_resolve_ip() will sleep in kzalloc() and wait_for_completion().
However, developers will not see any warnings if they use ib_init_ah_from_wc()
in an atomic context and test only on IB, because the function doesn't
sleep in that case.
Add a might_sleep() so that lockdep will catch bugs no matter what hardware is
used to test.
Signed-off-by: Roland Dreier <roland@purestorage.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
A couple of places in the CM do
spin_lock_irq(&cm_id_priv->lock);
...
if (cm_alloc_response_msg(work->port, work->mad_recv_wc, &msg))
However when the underlying transport is RoCE, this leads to a sleeping function
being called with the lock held - the callchain is
cm_alloc_response_msg() ->
ib_create_ah_from_wc() ->
ib_init_ah_from_wc() ->
rdma_addr_find_l2_eth_by_grh() ->
rdma_resolve_ip()
and rdma_resolve_ip() starts out by doing
req = kzalloc(sizeof *req, GFP_KERNEL);
not to mention rdma_addr_find_l2_eth_by_grh() doing
wait_for_completion(&ctx.comp);
to wait for the task that rdma_resolve_ip() queues up.
Fix this by moving the AH creation out of the lock.
Signed-off-by: Roland Dreier <roland@purestorage.com>
Reviewed-by: Sean Hefty <sean.hefty@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Minghuan Lian