To support disassociation and PCI hot unplug, we have to track all the
VMAs that refer to the device IO memory. When disassociation occurs the
VMAs have to be revised to point to the zero page, not the IO memory, to
allow the physical HW to be unplugged.
The three drivers supporting this implemented three different versions
of this algorithm, all leaving something to be desired. This new common
implementation has a few differences from the driver versions:
- Track all VMAs, including splitting/truncating/etc. Tie the lifetime of
the private data allocation to the lifetime of the vma. This avoids any
tricks with setting vm_ops which Linus didn't like. (see link)
- Support multiple mms, and support properly tracking mmaps triggered by
processes other than the one first opening the uverbs fd. This makes
fork behavior of disassociation enabled drivers the same as fork support
in normal drivers.
- Don't use crazy get_task stuff.
- Simplify the approach for to racing between vm_ops close and
disassociation, fixing the related bugs most of the driver
implementations had. Since we are in core code the tracking list can be
placed in struct ib_uverbs_ufile, which has a lifetime strictly longer
than any VMAs created by mmap on the uverbs FD.
Link: https://www.spinics.net/lists/stable/msg248747.html
Link: https://lkml.kernel.org/r/CA+55aFxJTV_g46AQPoPXen-UPiqR1HGMZictt7VpC-SMFbm3Cw@mail.gmail.com
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
Convert the ioctl method syscall path to use the uverbs_api data
structures. The new uapi structure includes all the same information, just
in a different and more optimal way.
- Use attr_bkey instead of 2 level radix trees for everything related to
attributes. This includes the attribute storage, presence, and
detection of missing mandatory attributes.
- Avoid iterating over all attribute storage at finish, instead use
find_first_bit with the attr_bkey to locate only those attrs that need
cleanup.
- Organize things to always run, and always rely on, cleanup. This
avoids a bunch of tricky error unwind cases.
- Locate the method using the radix tree, and locate the attributes
using a very efficient incremental radix tree lookup
- Use the precomputed destroy_bkey to handle uobject destruction
- Use the precomputed allocation sizes and precomputed 'need_stack'
to avoid maths in the fast path. This is optimal if userspace
does not pass (many) unsupported attributes.
Overall this results in much better codegen for the attribute accessors,
everything is now stored in bitmaps or linear arrays indexed by attr_bkey.
The compiler can compute attr_bkey values at compile time for all method
attributes, meaning things like uverbs_attr_is_valid() now compile into
single instruction bit tests.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Currently the struct uverbs_obj_type stored in the ib_uobject is part of
the .rodata segment of the module that defines the object. This is a
problem if drivers define new uapi objects as we will be left with a
dangling pointer after device disassociation.
Switch the uverbs_obj_type for struct uverbs_api_object, which is
allocated memory that is part of the uverbs_api and is guaranteed to
always exist. Further this moves the 'type_class' into this memory which
means access to the IDR/FD function pointers is also guaranteed. Drivers
cannot define new types.
This makes it safe to continue to use all uobjects, including driver
defined ones, after disassociation.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
This radix tree datastructure is intended to replace the 'hash' structure
used today for parsing ioctl methods during system calls. This first
commit introduces the structure and builds it from the existing .rodata
descriptions.
The so-called hash arrangement is actually a 5 level open coded radix tree.
This new version uses a 3 level radix tree built using the radix tree
library.
Overall this is much less code and much easier to build as the radix tree
API allows for dynamic modification during the building. There is a small
memory penalty to pay for this, but since the radix tree is allocated on
a per device basis, a few kb of RAM seems immaterial considering the
gained simplicity.
The radix tree is similar to the existing tree, but also has a 'attr_bkey'
concept, which is a small value'd index for each method attribute. This is
used to simplify and improve performance of everything in the next
patches.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
The disassociate function was broken by design because it failed all
commands. This prevents userspace from calling destroy on a uobject after
it has detected a device fatal error and thus reclaiming the resources in
userspace is prevented.
This fix is now straightforward, when anything destroys a uobject that is
not the user the object remains on the IDR with a NULL context and object
pointer. All lookup locking modes other than DESTROY will fail. When the
user ultimately calls the destroy function it is simply dropped from the
IDR while any related information is returned.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
After all the recent structural changes this is now straightfoward, hoist
the hw_destroy_rwsem up out of rdma_destroy_explicit and wrap it around
the uobject write lock as well as the destroy.
This is necessary as obtaining a write lock concurrently with
uverbs_destroy_ufile_hw() will cause malfunction.
After this change none of the destroy callbacks require the
disassociate_srcu lock to be correct.
This requires introducing a new lookup mode, UVERBS_LOOKUP_DESTROY as the
IOCTL interface needs to hold an unlocked kref until all command
verification is completed.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The locking here has always been a bit crazy and spread out, upon some
careful analysis we can simplify things.
Create a single function uverbs_destroy_ufile_hw() that internally handles
all locking. This pulls together pieces of this process that were
sprinkled all over the places into one place, and covers them with one
lock.
This eliminates several duplicate/confusing locks and makes the control
flow in ib_uverbs_close() and ib_uverbs_free_hw_resources() extremely
simple.
Unfortunately we have to keep an extra mutex, ucontext_lock. This lock is
logically part of the rwsem and provides the 'down write, fail if write
locked, wait if read locked' semantic we require.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Our ABI for write() uses a s32 for FDs and a u32 for IDRs, but internally
we ended up implicitly casting these ABI values into an 'int'. For ioctl()
we use a s64 for FDs and a u64 for IDRs, again casting to an int.
The various casts to int are all missing range checks which can cause
userspace values that should be considered invalid to be accepted.
Fix this by making the generic lookup routine accept a s64, which does not
truncate the write API's u32/s32 or the ioctl API's s64. Then push the
detailed range checking down to the actual type implementations to be
shared by both interfaces.
Finally, change the copy of the uobj->id to sign extend into a s64, so eg,
if we ever wish to return a negative value for a FD it is carried
properly.
This ensures that userspace values are never weirdly interpreted due to
the various trunctations and everything that is really out of range gets
an EINVAL.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
The correct handle to refer to the idr/etc is ib_uverbs_file, revise all
the core APIs to use this instead. The user API are left as wrappers
that automatically convert a ucontext to a ufile for now.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
The IDR is part of the ib_ufile so all the machinery to lock it, handle
closing and disassociation rightly belongs to the ufile not the ucontext.
This changes the lifetime of that data to match the lifetime of the file
descriptor which is always strictly longer than the lifetime of the
ucontext.
We need the entire locking machinery to continue to exist after ucontext
destruction to allow us to return the destroy data after a device has been
disassociated.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
The specs are required to operate the uverbs file, so they belong inside
the ib_uverbs_device, not inside the ib_device. The spec passed in the
ib_device is just a communication from the driver and should not be used
during runtime.
This also changes the lifetime of the spec memory to match the
ib_uverbs_device, however at this time the spec_root can still contain
driver pointers after disassociation, so it cannot be used if ib_dev is
NULL. This is preparation for another series.
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
uverbs_finalize_objects is currently used only to commit or abort
objects. Since we want to add automatic allocation/free of PTR_IN
attributes, moving it to uverbs_ioctl.c and renamit it to
uverbs_finalize_attrs.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.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>
The new ioctl based infrastructure either commits or rollbacks
all objects of the method as one transaction. In order to do
that, we introduce a notion of dealing with a collection of
objects that are related to a specific method.
This also requires adding a notion of a method and attribute.
A method contains a hash of attributes, where each bucket
contains several attributes. The attributes are hashed according
to their namespace which resides in the four upper bits of the id.
For example, an object could be a CQ, which has an action of CREATE_CQ.
This action has multiple attributes. For example, the CQ's new handle
and the comp_channel. Each layer in this hierarchy - objects, methods
and attributes is split into namespaces. The basic example for that is
one namespace representing the default entities and another one
representing the driver specific entities.
When declaring these methods and attributes, we actually declare
their specifications. When a method is executed, we actually
allocates some space to hold auxiliary information. This auxiliary
information contains meta-data about the required objects, such
as pointers to their type information, pointers to the uobjects
themselves (if exist), etc.
The specification, along with the auxiliary information we allocated
and filled is given to the finalize_objects function.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The ioctl infrastructure treats all user-objects in the same manner.
It gets objects ids from the user-space and by using the object type
and type attributes mentioned in the object specification, it executes
this required method. Passing an object id from the user-space as
an attribute is carried out in three stages. The first is carried out
before the actual handler and the last is carried out afterwards.
The different supported operations are read, write, destroy and create.
In the first stage, the former three actions just fetches the object
from the repository (by using its id) and locks it. The last action
allocates a new uobject. Afterwards, the second stage is carried out
when the handler itself carries out the required modification of the
object. The last stage is carried out after the handler finishes and
commits the result. The former two operations just unlock the object.
Destroy calls the "free object" operation, taking into account the
object's type and releases the uobject as well. Creation just adds the
new uobject to the repository, making the object visible to the
application.
In order to abstract these details from the ioctl infrastructure
layer, we add uverbs_get_uobject_from_context and
uverbs_finalize_object functions which corresponds to the first
and last stages respectively.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The completion channel we use in verbs infrastructure is FD based.
Previously, we had a separate way to manage this object. Since we
strive for a single way to manage any kind of object in this
infrastructure, we conceptually treat all objects as subclasses
of ib_uobject.
This commit adds the necessary mechanism to support FD based objects
like their IDR counterparts. FD objects release need to be synchronized
with context release. We use the cleanup_mutex on the uverbs_file for
that.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
This changes only the handlers which deals with idr based objects to
use the new idr allocation, fetching and destruction schema.
This patch consists of the following changes:
(1) Allocation, fetching and destruction is done via idr ops.
(2) Context initializing and release is done through
uverbs_initialize_ucontext and uverbs_cleanup_ucontext.
(3) Ditching the live flag. Mostly, this is pretty straight
forward. The only place that is a bit trickier is in
ib_uverbs_open_qp. Commit [1] added code to check whether
the uobject is already live and initialized. This mostly
happens because of a race between open_qp and events.
We delayed assigning the uobject's pointer in order to
eliminate this race without using the live variable.
[1] commit a040f95dc8
("IB/core: Fix XRC race condition in ib_uverbs_open_qp")
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
The new ioctl infrastructure supports driver specific objects.
Each such object type has a hot unplug function, allocation size and
an order of destruction.
When a ucontext is created, a new list is created in this ib_ucontext.
This list contains all objects created under this ib_ucontext.
When a ib_ucontext is destroyed, we traverse this list several time
destroying the various objects by the order mentioned in the object
type description. If few object types have the same destruction order,
they are destroyed in an order opposite to their creation.
Adding an object is done in two parts.
First, an object is allocated and added to idr tree. Then, the
command's handlers (in downstream patches) could work on this object
and fill in its required details.
After a successful command, the commit part is called and the user
objects become ucontext visible. If the handler failed, alloc_abort
should be called.
Removing an uboject is done by calling lookup_get with the write flag
and finalizing it with destroy_commit. A major change from the previous
code is that we actually destroy the kernel object itself in
destroy_commit (rather than just the uobject).
We should make sure idr (per-uverbs-file) and list (per-ucontext) could
be accessed concurrently without corrupting them.
Signed-off-by: Matan Barak <matanb@mellanox.com>
Reviewed-by: Yishai Hadas <yishaih@mellanox.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>
