Under nested IOMMU translation, userspace owns the stage-1 translation
table (e.g. the stage-1 page table of Intel VT-d or the context table of
ARM SMMUv3, and etc.). Stage-1 translation tables are vendor specific, and
need to be compatible with the underlying IOMMU hardware. Hence, userspace
should know the IOMMU hardware capability before creating and configuring
the stage-1 translation table to kernel.
This adds IOMMU_GET_HW_INFO ioctl to query the IOMMU hardware information
(a.k.a capability) for a given device. The returned data is vendor
specific, userspace needs to decode it with the structure by the output
@out_data_type field.
As only physical devices have IOMMU hardware, so this will return error if
the given device is not a physical device.
Link: https://lore.kernel.org/r/20230818101033.4100-4-yi.l.liu@intel.com
Reviewed-by: Lu Baolu <baolu.lu@linux.intel.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Co-developed-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Yi Liu <yi.l.liu@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Required for following patches.
Resolve merge conflict by using the hunk from the for-next branch and
shifting the iommufd_object_deref_user() into iommufd_hw_pagetable_put()
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The complication of the mutex and refcount will be amplified after we
introduce the replace support for accesses. So, add a preparatory change
of a constitutive helper iommufd_access_change_ioas() and its wrapper
iommufd_access_change_ioas_id(). They can simply take care of existing
iommufd_access_attach() and iommufd_access_detach(), properly sequencing
the refcount puts so that they are truely at the end of the sequence after
we know the IOAS pointer is not required any more.
Link: https://lore.kernel.org/r/da0c462532193b447329c4eb975a596f47e49b70.1690523699.git.nicolinc@nvidia.com
Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This is a preparatory change for ioas replacement support for accesses.
The replacement routine does an iopt_add_access() for a new IOAS first and
then iopt_remove_access() for the old IOAS upon the success of the first
call. However, the first call overrides the iopt_access_list_id in the
access struct, resulting in iopt_remove_access() being unable to work on
the old IOAS.
Add an iopt_access_list_id as a parameter to iopt_remove_access, so the
replacement routine can save the id before it gets overwritten. Pass the
id in iopt_remove_access() for a proper cleanup.
The existing callers should just pass in access->iopt_access_list_id.
Link: https://lore.kernel.org/r/7bb939b9e0102da0c099572bb3de78ab7622221e.1690523699.git.nicolinc@nvidia.com
Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
syzkaller found a race where IOMMUFD_DESTROY increments the refcount:
obj = iommufd_get_object(ucmd->ictx, cmd->id, IOMMUFD_OBJ_ANY);
if (IS_ERR(obj))
return PTR_ERR(obj);
iommufd_ref_to_users(obj);
/* See iommufd_ref_to_users() */
if (!iommufd_object_destroy_user(ucmd->ictx, obj))
As part of the sequence to join the two existing primitives together.
Allowing the refcount the be elevated without holding the destroy_rwsem
violates the assumption that all temporary refcount elevations are
protected by destroy_rwsem. Racing IOMMUFD_DESTROY with
iommufd_object_destroy_user() will cause spurious failures:
WARNING: CPU: 0 PID: 3076 at drivers/iommu/iommufd/device.c:477 iommufd_access_destroy+0x18/0x20 drivers/iommu/iommufd/device.c:478
Modules linked in:
CPU: 0 PID: 3076 Comm: syz-executor.0 Not tainted 6.3.0-rc1-syzkaller #0
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 07/03/2023
RIP: 0010:iommufd_access_destroy+0x18/0x20 drivers/iommu/iommufd/device.c:477
Code: e8 3d 4e 00 00 84 c0 74 01 c3 0f 0b c3 0f 1f 44 00 00 f3 0f 1e fa 48 89 fe 48 8b bf a8 00 00 00 e8 1d 4e 00 00 84 c0 74 01 c3 <0f> 0b c3 0f 1f 44 00 00 41 57 41 56 41 55 4c 8d ae d0 00 00 00 41
RSP: 0018:ffffc90003067e08 EFLAGS: 00010246
RAX: 0000000000000000 RBX: ffff888109ea0300 RCX: 0000000000000000
RDX: 0000000000000001 RSI: 0000000000000000 RDI: 00000000ffffffff
RBP: 0000000000000004 R08: 0000000000000000 R09: ffff88810bbb3500
R10: ffff88810bbb3e48 R11: 0000000000000000 R12: ffffc90003067e88
R13: ffffc90003067ea8 R14: ffff888101249800 R15: 00000000fffffffe
FS: 00007ff7254fe6c0(0000) GS:ffff888237c00000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000555557262da8 CR3: 000000010a6fd000 CR4: 0000000000350ef0
Call Trace:
<TASK>
iommufd_test_create_access drivers/iommu/iommufd/selftest.c:596 [inline]
iommufd_test+0x71c/0xcf0 drivers/iommu/iommufd/selftest.c:813
iommufd_fops_ioctl+0x10f/0x1b0 drivers/iommu/iommufd/main.c:337
vfs_ioctl fs/ioctl.c:51 [inline]
__do_sys_ioctl fs/ioctl.c:870 [inline]
__se_sys_ioctl fs/ioctl.c:856 [inline]
__x64_sys_ioctl+0x84/0xc0 fs/ioctl.c:856
do_syscall_x64 arch/x86/entry/common.c:50 [inline]
do_syscall_64+0x38/0x80 arch/x86/entry/common.c:80
entry_SYSCALL_64_after_hwframe+0x63/0xcd
The solution is to not increment the refcount on the IOMMUFD_DESTROY path
at all. Instead use the xa_lock to serialize everything. The refcount
check == 1 and xa_erase can be done under a single critical region. This
avoids the need for any refcount incrementing.
It has the downside that if userspace races destroy with other operations
it will get an EBUSY instead of waiting, but this is kind of racing is
already dangerous.
Fixes: 2ff4bed7fe ("iommufd: File descriptor, context, kconfig and makefiles")
Link: https://lore.kernel.org/r/2-v1-85aacb2af554+bc-iommufd_syz3_jgg@nvidia.com
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Reported-by: syzbot+7574ebfe589049630608@syzkaller.appspotmail.com
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Replace allows all the devices in a group to move in one step to a new
HWPT. Further, the HWPT move is done without going through a blocking
domain so that the IOMMU driver can implement some level of
non-distruption to ongoing DMA if that has meaning for it (eg for future
special driver domains)
Replace uses a lot of the same logic as normal attach, except the actual
domain change over has different restrictions, and we are careful to
sequence things so that failure is going to leave everything the way it
was, and not get trapped in a blocking domain or something if there is
ENOMEM.
Link: https://lore.kernel.org/r/14-v8-6659224517ea+532-iommufd_alloc_jgg@nvidia.com
Reviewed-by: Lu Baolu <baolu.lu@linux.intel.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Tested-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The code flow for first time attaching a PT and replacing a PT is very
similar except for the lowest do_attach step.
Reorganize this so that the do_attach step is a function pointer.
Replace requires destroying the old HWPT once it is replaced. This
destruction cannot be done under all the locks that are held in the
function pointer, so the signature allows returning a HWPT which will be
destroyed by the caller after everything is unlocked.
Link: https://lore.kernel.org/r/12-v8-6659224517ea+532-iommufd_alloc_jgg@nvidia.com
Reviewed-by: Lu Baolu <baolu.lu@linux.intel.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Tested-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Due to the auto_domains mechanism the ioas->mutex must be held until
the hwpt is completely setup by iommufd_object_abort_and_destroy() or
iommufd_object_finalize().
This prevents a concurrent iommufd_device_auto_get_domain() from seeing
an incompletely initialized object through the ioas->hwpt_list.
To make this more consistent move the unlock until after finalize.
Fixes: e8d5721003 ("iommufd: Add kAPI toward external drivers for physical devices")
Link: https://lore.kernel.org/r/11-v8-6659224517ea+532-iommufd_alloc_jgg@nvidia.com
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Tested-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The sw_msi_start is only set by the ARM drivers and it is always constant.
Due to the way vfio/iommufd allow domains to be re-used between
devices we have a built in assumption that there is only one value
for sw_msi_start and it is global to the system.
To make replace simpler where we may not reparse the
iommu_get_resv_regions() move the sw_msi_start to the iommufd_group so it
is always available once any HWPT has been attached.
Link: https://lore.kernel.org/r/7-v8-6659224517ea+532-iommufd_alloc_jgg@nvidia.com
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Tested-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The driver facing API in the iommu core makes the reserved regions
per-device. An algorithm in the core code consolidates the regions of all
the devices in a group to return the group view.
To allow for devices to be hotplugged into the group iommufd would re-load
the entire group's reserved regions for each device, just in case they
changed.
Further iommufd already has to deal with duplicated/overlapping reserved
regions as it must union all the groups together.
Thus simplify all of this to just use the device reserved regions
interface directly from the iommu driver.
Link: https://lore.kernel.org/r/5-v8-6659224517ea+532-iommufd_alloc_jgg@nvidia.com
Suggested-by: Kevin Tian <kevin.tian@intel.com>
Reviewed-by: Lu Baolu <baolu.lu@linux.intel.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Tested-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The devices list was used as a simple way to avoid having per-group
information. Now that this seems to be unavoidable, just commit to
per-group information fully and remove the devices list from the HWPT.
The iommufd_group stores the currently assigned HWPT for the entire group
and we can manage the per-device attach/detach with a list in the
iommufd_group.
For destruction the flow is organized to make the following patches
easier, the actual call to iommufd_object_destroy_user() is done at the
top of the call chain without holding any locks. The HWPT to be destroyed
is returned out from the locked region to make this possible. Later
patches create locking that requires this.
Link: https://lore.kernel.org/r/3-v8-6659224517ea+532-iommufd_alloc_jgg@nvidia.com
Reviewed-by: Lu Baolu <baolu.lu@linux.intel.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Tested-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Previously, the detach routine is only done by the destroy(). And it was
called by vfio_iommufd_emulated_unbind() when the device runs close(), so
all the mappings in iopt were cleaned in that setup, when the call trace
reaches this detach() routine.
Now, there's a need of a detach uAPI, meaning that it does not only need
a new iommufd_access_detach() API, but also requires access->ops->unmap()
call as a cleanup. So add one.
However, leaving that unprotected can introduce some potential of a race
condition during the pin_/unpin_pages() call, where access->ioas->iopt is
getting referenced. So, add an ioas_lock to protect the context of iopt
referencings.
Also, to allow the iommufd_access_unpin_pages() callback to happen via
this unmap() call, add an ioas_unpin pointer, so the unpin routine won't
be affected by the "access->ioas = NULL" trick.
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Tested-by: Terrence Xu <terrence.xu@intel.com>
Tested-by: Nicolin Chen <nicolinc@nvidia.com>
Tested-by: Matthew Rosato <mjrosato@linux.ibm.com>
Tested-by: Yanting Jiang <yanting.jiang@intel.com>
Tested-by: Shameer Kolothum <shameerali.kolothum.thodi@huawei.com>
Tested-by: Zhenzhong Duan <zhenzhong.duan@intel.com>
Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Yi Liu <yi.l.liu@intel.com>
Link: https://lore.kernel.org/r/20230718135551.6592-15-yi.l.liu@intel.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Yi Liu says
===================
The .bind_iommufd op of vfio emulated devices are either empty or does
nothing. This is different with the vfio physical devices, to add vfio
device cdev, need to make them act the same.
This series first makes the .bind_iommufd op of vfio emulated devices to
create iommufd_access, this introduces a new iommufd API. Then let the
driver that does not provide .bind_iommufd op to use the vfio emulated
iommufd op set. This makes all vfio device drivers have consistent iommufd
operations, which is good for adding new device uAPIs in the device cdev
===================
* branch 'vfio_mdev_ops':
vfio: Check the presence for iommufd callbacks in __vfio_register_dev()
vfio/mdev: Uses the vfio emulated iommufd ops set in the mdev sample drivers
vfio-iommufd: Make vfio_iommufd_emulated_bind() return iommufd_access ID
vfio-iommufd: No need to record iommufd_ctx in vfio_device
iommufd: Create access in vfio_iommufd_emulated_bind()
iommu/iommufd: Pass iommufd_ctx pointer in iommufd_get_ioas()
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
There are needs to created iommufd_access prior to have an IOAS and set
IOAS later. Like the vfio device cdev needs to have an iommufd object
to represent the bond (iommufd_access) and IOAS replacement.
Moves the iommufd_access_create() call into vfio_iommufd_emulated_bind(),
making it symmetric with the __vfio_iommufd_access_destroy() call in the
vfio_iommufd_emulated_unbind(). This means an access is created/destroyed
by the bind()/unbind(), and the vfio_iommufd_emulated_attach_ioas() only
updates the access->ioas pointer.
Since vfio_iommufd_emulated_bind() does not provide ioas_id, drop it from
the argument list of iommufd_access_create(). Instead, add a new access
API iommufd_access_attach() to set the access->ioas pointer. Also, set
vdev->iommufd_attached accordingly, similar to the physical pathway.
Link: https://lore.kernel.org/r/20230327093351.44505-3-yi.l.liu@intel.com
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Tested-by: Terrence Xu <terrence.xu@intel.com>
Tested-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
Signed-off-by: Yi Liu <yi.l.liu@intel.com>
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
iommufd wants to use more infrastructure, like the iommu_group, that the
mock device does not support. Create a more complete mock device that can
go through the whole cycle of ownership, blocking domain, and has an
iommu_group.
This requires creating a real struct device on a real bus to be able to
connect it to a iommu_group. Unfortunately we cannot formally attach the
mock iommu driver as an actual driver as the iommu core does not allow
more than one driver or provide a general way for busses to link to
iommus. This can be solved with a little hack to open code the dev_iommus
struct.
With this infrastructure things work exactly the same as the normal domain
path, including the auto domains mechanism and direct attach of hwpts. As
the created hwpt is now an autodomain it is no longer required to destroy
it and trying to do so will trigger a failure.
Link: https://lore.kernel.org/r/11-v3-ae9c2975a131+2e1e8-iommufd_hwpt_jgg@nvidia.com
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The HWPT is always linked to an IOAS and once a HWPT exists its domain
should be fully mapped. This ended up being split up into device.c during
a two phase creation that was a bit confusing.
Move the iopt_table_add_domain() into iommufd_hw_pagetable_alloc() by
having it call back to device.c to complete the domain attach in the
required order.
Calling iommufd_hw_pagetable_alloc() with immediate_attach = false will
work on most drivers, but notably the SMMU drivers will fail because they
can't decide what kind of domain to create until they are attached. This
will be fixed when the domain_alloc function can take in a struct device.
Link: https://lore.kernel.org/r/6-v3-ae9c2975a131+2e1e8-iommufd_hwpt_jgg@nvidia.com
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
A HWPT is permanently associated with an IOAS when it is created, remove
the strange situation where a refcount != 0 HWPT can have been
disconnected from the IOAS by putting all the IOAS related destruction in
the object destroy function.
Initializing a HWPT is two stages, we have to allocate it, attach it to a
device and then populate the domain. Once the domain is populated it is
fully linked to the IOAS.
Arrange things so that all the error unwinds flow through the
iommufd_hw_pagetable_destroy() and allow it to handle all cases.
Link: https://lore.kernel.org/r/4-v3-ae9c2975a131+2e1e8-iommufd_hwpt_jgg@nvidia.com
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
The hwpt is added to the hwpt_list only during its creation, it is never
added again. This hunk is some missed leftover from rework. Adding it
twice will corrupt the linked list in some cases.
It effects HWPT specific attachment, which is something the test suite
cannot cover until we can create a legitimate struct device with a
non-system iommu "driver" (ie we need the bus removed from the iommu code)
Cc: stable@vger.kernel.org
Fixes: e8d5721003 ("iommufd: Add kAPI toward external drivers for physical devices")
Link: https://lore.kernel.org/r/1-v1-4336b5cb2fe4+1d7-iommufd_hwpt_jgg@nvidia.com
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Reported-by: Kevin Tian <kevin.tian@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Eric points out this is wrong for the rare case of someone using
allow_unsafe_interrupts on ARM. We always have to setup the MSI window in
the domain if the iommu driver asks for it.
Move the iommu_get_msi_cookie() setup to the top of the function and
always do it, regardless of the security mode. Add checks to
iommufd_device_setup_msi() to ensure the driver is not doing something
incomprehensible. No current driver will set both a HW and SW MSI window,
or have more than one SW MSI window.
Fixes: e8d5721003 ("iommufd: Add kAPI toward external drivers for physical devices")
Link: https://lore.kernel.org/r/3-v1-0362a1a1c034+98-iommufd_fixes1_jgg@nvidia.com
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Reported-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Provide a mock kernel module for the iommu_domain that allows it to run
without any HW and the mocking provides a way to directly validate that
the PFNs loaded into the iommu_domain are correct. This exposes the access
kAPI toward userspace to allow userspace to explore the functionality of
pages.c and io_pagetable.c
The mock also simulates the rare case of PAGE_SIZE > iommu page size as
the mock will operate at a 2K iommu page size. This allows exercising all
of the calculations to support this mismatch.
This is also intended to support syzkaller exploring the same space.
However, it is an unusually invasive config option to enable all of
this. The config option should not be enabled in a production kernel.
Link: https://lore.kernel.org/r/16-v6-a196d26f289e+11787-iommufd_jgg@nvidia.com
Tested-by: Matthew Rosato <mjrosato@linux.ibm.com> # s390
Tested-by: Eric Auger <eric.auger@redhat.com> # aarch64
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Kernel access is the mode that VFIO "mdevs" use. In this case there is no
struct device and no IOMMU connection. iommufd acts as a record keeper for
accesses and returns the actual struct pages back to the caller to use
however they need. eg with kmap or the DMA API.
Each caller must create a struct iommufd_access with
iommufd_access_create(), similar to how iommufd_device_bind() works. Using
this struct the caller can access blocks of IOVA using
iommufd_access_pin_pages() or iommufd_access_rw().
Callers must provide a callback that immediately unpins any IOVA being
used within a range. This happens if userspace unmaps the IOVA under the
pin.
The implementation forwards the access requests directly to the iopt
infrastructure that manages the iopt_pages_access.
Link: https://lore.kernel.org/r/14-v6-a196d26f289e+11787-iommufd_jgg@nvidia.com
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Tested-by: Nicolin Chen <nicolinc@nvidia.com>
Tested-by: Yi Liu <yi.l.liu@intel.com>
Tested-by: Lixiao Yang <lixiao.yang@intel.com>
Tested-by: Matthew Rosato <mjrosato@linux.ibm.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Add the four functions external drivers need to connect physical DMA to
the IOMMUFD:
iommufd_device_bind() / iommufd_device_unbind()
Register the device with iommufd and establish security isolation.
iommufd_device_attach() / iommufd_device_detach()
Connect a bound device to a page table
Binding a device creates a device object ID in the uAPI, however the
generic API does not yet provide any IOCTLs to manipulate them.
Link: https://lore.kernel.org/r/13-v6-a196d26f289e+11787-iommufd_jgg@nvidia.com
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Tested-by: Nicolin Chen <nicolinc@nvidia.com>
Tested-by: Yi Liu <yi.l.liu@intel.com>
Tested-by: Lixiao Yang <lixiao.yang@intel.com>
Tested-by: Matthew Rosato <mjrosato@linux.ibm.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
