MAP_DMA ioctls might be called from various threads within a process,
for example when using QEMU, the vCPU threads are often generating
these calls and we therefore take a reference to that vCPU task.
However, QEMU also supports vCPU hotplug on some machines and the task
that called MAP_DMA may have exited by the time UNMAP_DMA is called,
resulting in the mm_struct pointer being NULL and thus a failure to
match against the existing mapping.
To resolve this, we instead take a reference to the thread
group_leader, which has the same mm_struct and resource limits, but
is less likely exit, at least in the QEMU case. A difficulty here is
guaranteeing that the capabilities of the group_leader match that of
the calling thread, which we resolve by tracking CAP_IPC_LOCK at the
time of calling rather than at an indeterminate time in the future.
Potentially this also results in better efficiency as this is now
recorded once per MAP_DMA ioctl.
Reported-by: Xu Yandong <xuyandong2@huawei.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Bisection by Amadeusz Sławiński implicates this commit leading to bad
page state issues after VM shutdown, likely due to unbalanced page
references. The original commit was intended only as a performance
improvement, therefore revert for offline rework.
Link: https://lkml.org/lkml/2018/6/2/97
Fixes: 356e88ebe4 ("vfio/type1: Improve memory pinning process for raw PFN mapping")
Cc: Jason Cai (Xiang Feng) <jason.cai@linux.alibaba.com>
Reported-by: Amadeusz Sławiński <amade@asmblr.net>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
When using vfio to pass through a PCIe device (e.g. a GPU card) that
has a huge BAR (e.g. 16GB), a lot of cycles are wasted on memory
pinning because PFNs of PCI BAR are not backed by struct page, and
the corresponding VMA has flag VM_PFNMAP.
With this change, when pinning a region which is a raw PFN mapping,
it can skip unnecessary user memory pinning process, and thus, can
significantly improve VM's boot up time when passing through devices
via VFIO. In my test on a Xeon E5 2.6GHz, the time mapping a 16GB
BAR was reduced from about 0.4s to 1.5us.
Signed-off-by: Jason Cai (Xiang Feng) <jason.cai@linux.alibaba.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
VFIO IOMMU type1 currently upmaps IOVA pages synchronously, which requires
IOTLB flushing for every unmapping. This results in large IOTLB flushing
overhead when handling pass-through devices has a large number of mapped
IOVAs. This can be avoided by using the new IOTLB flushing interface.
Cc: Alex Williamson <alex.williamson@redhat.com>
Cc: Joerg Roedel <joro@8bytes.org>
Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
[aw - use LIST_HEAD]
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Filesystem-DAX is incompatible with 'longterm' page pinning. Without
page cache indirection a DAX mapping maps filesystem blocks directly.
This means that the filesystem must not modify a file's block map while
any page in a mapping is pinned. In order to prevent the situation of
userspace holding of filesystem operations indefinitely, disallow
'longterm' Filesystem-DAX mappings.
RDMA has the same conflict and the plan there is to add a 'with lease'
mechanism to allow the kernel to notify userspace that the mapping is
being torn down for block-map maintenance. Perhaps something similar can
be put in place for vfio.
Note that xfs and ext4 still report:
"DAX enabled. Warning: EXPERIMENTAL, use at your own risk"
...at mount time, and resolving the dax-dma-vs-truncate problem is one
of the last hurdles to remove that designation.
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: kvm@vger.kernel.org
Cc: <stable@vger.kernel.org>
Reported-by: Haozhong Zhang <haozhong.zhang@intel.com>
Tested-by: Haozhong Zhang <haozhong.zhang@intel.com>
Fixes: d475c6346a ("dax,ext2: replace XIP read and write with DAX I/O")
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
I get a static checker warning about the potential integer overflow if
we add "unmap->iova + unmap->size". The integer overflow isn't really
harmful, but we may as well fix it. Also unmap->size gets truncated to
size_t when we pass it to vfio_find_dma() so we could check for too high
values of that as well.
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
If the IOMMU driver advertises 'real' reserved regions for MSIs, but
still includes the software-managed region as well, we are currently
blind to the former and will configure the IOMMU domain to map MSIs into
the latter, which is unlikely to work as expected.
Since it would take a ridiculous hardware topology for both regions to
be valid (which would be rather difficult to support in general), we
should be safe to assume that the presence of any hardware regions makes
the software region irrelevant. However, the IOMMU driver might still
advertise the software region by default, particularly if the hardware
regions are filled in elsewhere by generic code, so it might not be fair
for VFIO to be super-strict about not mixing them. To that end, make
vfio_iommu_has_sw_msi() robust against the presence of both region types
at once, so that we end up doing what is almost certainly right, rather
than what is almost certainly wrong.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Tested-by: Shameer Kolothum <shameerali.kolothum.thodi@huawei.com>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
For ARM-based systems with a GICv3 ITS to provide interrupt isolation,
but hardware limitations which are worked around by having MSIs bypass
SMMU translation (e.g. HiSilicon Hip06/Hip07), VFIO neglects to check
for the IRQ_DOMAIN_FLAG_MSI_REMAP capability, (and thus erroneously
demands unsafe_interrupts) if a software-managed MSI region is absent.
Fix this by always checking for isolation capability at both the IRQ
domain and IOMMU domain levels, rather than predicating that on whether
MSIs require an IOMMU mapping (which was always slightly tenuous logic).
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Tested-by: Shameer Kolothum <shameerali.kolothum.thodi@huawei.com>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
vfio_pin_pages_remote() is typically called to iterate over a range
of memory. Testing CAP_IPC_LOCK is relatively expensive, so it makes
sense to push it up to the caller, which can then repeatedly call
vfio_pin_pages_remote() using that value. This can show nearly a 20%
improvement on the worst case path through VFIO_IOMMU_MAP_DMA with
contiguous page mapping disabled. Testing RLIMIT_MEMLOCK is much more
lightweight, but we bring it along on the same principle and it does
seem to show a marginal improvement.
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
With vfio_lock_acct() testing the locked memory limit under mmap_sem,
it's redundant to do it here for a single page. We can also reorder
our tests such that we can avoid testing for reserved pages if we're
not doing accounting and let vfio_lock_acct() test the process
CAP_IPC_LOCK. Finally, this function oddly returns 1 on success.
Update to return zero on success, -errno on error. Since the function
only pins a single page, there's no need to return the number of pages
pinned.
N.B. vfio_pin_pages_remote() can pin a large contiguous range of pages
before calling vfio_lock_acct(). If we were to similarly remove the
extra test there, a user could temporarily pin far more pages than
they're allowed.
Suggested-by: Kirti Wankhede <kwankhede@nvidia.com>
Suggested-by: Eric Auger <eric.auger@redhat.com>
Reviewed-by: Kirti Wankhede <kwankhede@nvidia.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
If the mmap_sem is contented then the vfio type1 IOMMU backend will
defer locked page accounting updates to a workqueue task. This has a
few problems and depending on which side the user tries to play, they
might be over-penalized for unmaps that haven't yet been accounted or
race the workqueue to enter more mappings than they're allowed. The
original intent of this workqueue mechanism seems to be focused on
reducing latency through the ioctl, but we cannot do so at the cost
of correctness. Remove this workqueue mechanism and update the
callers to allow for failure. We can also now recheck the limit under
write lock to make sure we don't exceed it.
vfio_pin_pages_remote() also now necessarily includes an unwind path
which we can jump to directly if the consecutive page pinning finds
that we're exceeding the user's memory limits. This avoids the
current lazy approach which does accounting and mapping up to the
fault, only to return an error on the next iteration to unwind the
entire vfio_dma.
Cc: stable@vger.kernel.org
Reviewed-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
The introduction of reserved regions has left a couple of rough edges
which we could do with sorting out sooner rather than later. Since we
are not yet addressing the potential dynamic aspect of software-managed
reservations and presenting them at arbitrary fixed addresses, it is
incongruous that we end up displaying hardware vs. software-managed MSI
regions to userspace differently, especially since ARM-based systems may
actually require one or the other, or even potentially both at once,
(which iommu-dma currently has no hope of dealing with at all). Let's
resolve the former user-visible inconsistency ASAP before the ABI has
been baked into a kernel release, in a way that also lays the groundwork
for the latter shortcoming to be addressed by follow-up patches.
For clarity, rename the software-managed type to IOMMU_RESV_SW_MSI, use
IOMMU_RESV_MSI to describe the hardware type, and document everything a
little bit. Since the x86 MSI remapping hardware falls squarely under
this meaning of IOMMU_RESV_MSI, apply that type to their regions as well,
so that we tell the same story to userspace across all platforms.
Secondly, as the various region types require quite different handling,
and it really makes little sense to ever try combining them, convert the
bitfield-esque #defines to a plain enum in the process before anyone
gets the wrong impression.
Fixes: d30ddcaa7b ("iommu: Add a new type field in iommu_resv_region")
Reviewed-by: Eric Auger <eric.auger@redhat.com>
CC: Alex Williamson <alex.williamson@redhat.com>
CC: David Woodhouse <dwmw2@infradead.org>
CC: kvm@vger.kernel.org
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
We are going to split <linux/sched/signal.h> out of <linux/sched.h>, which
will have to be picked up from other headers and a couple of .c files.
Create a trivial placeholder <linux/sched/signal.h> file that just
maps to <linux/sched.h> to make this patch obviously correct and
bisectable.
Include the new header in the files that are going to need it.
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
We are going to split <linux/sched/mm.h> out of <linux/sched.h>, which
will have to be picked up from other headers and a couple of .c files.
Create a trivial placeholder <linux/sched/mm.h> file that just
maps to <linux/sched.h> to make this patch obviously correct and
bisectable.
The APIs that are going to be moved first are:
mm_alloc()
__mmdrop()
mmdrop()
mmdrop_async_fn()
mmdrop_async()
mmget_not_zero()
mmput()
mmput_async()
get_task_mm()
mm_access()
mm_release()
Include the new header in the files that are going to need it.
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-kernel@vger.kernel.org
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Fix to return a negative error code from the error handling
case instead of 0, as done elsewhere in this function.
Fixes: 5d70499218 ("vfio/type1: Allow transparent MSI IOVA allocation")
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
In case the IOMMU translates MSI transactions (typical case
on ARM), we check MSI remapping capability at IRQ domain
level. Otherwise it is checked at IOMMU level.
At this stage the arm-smmu-(v3) still advertise the
IOMMU_CAP_INTR_REMAP capability at IOMMU level. This will be
removed in subsequent patches.
Signed-off-by: Eric Auger <eric.auger@redhat.com>
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Reviewed-by: Tomasz Nowicki <tomasz.nowicki@caviumnetworks.com>
Tested-by: Tomasz Nowicki <tomasz.nowicki@caviumnetworks.com>
Tested-by: Bharat Bhushan <bharat.bhushan@nxp.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
When attaching a group to the container, check the group's
reserved regions and test whether the IOMMU translates MSI
transactions. If yes, we initialize an IOVA allocator through
the iommu_get_msi_cookie API. This will allow the MSI IOVAs
to be transparently allocated on MSI controller's compose().
Signed-off-by: Eric Auger <eric.auger@redhat.com>
Acked-by: Alex Williamson <alex.williamson@redhat.com>
Reviewed-by: Tomasz Nowicki <tomasz.nowicki@caviumnetworks.com>
Tested-by: Tomasz Nowicki <tomasz.nowicki@caviumnetworks.com>
Tested-by: Bharat Bhushan <bharat.bhushan@nxp.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Using has_capability() rather than ns_capable(), we're no longer using
this header.
Cc: Jike Song <jike.song@intel.com>
Cc: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Before the mdev enhancement type1 iommu used capable() to test the
capability of current task; in the course of mdev development a
new requirement, testing for another task other than current, was
raised. ns_capable() was used for this purpose, however it still
tests current, the only difference is, in a specified namespace.
Fix it by using has_capability() instead, which tests the cap for
specified task in init_user_ns, the same namespace as capable().
Cc: Gerd Hoffmann <kraxel@redhat.com>
Signed-off-by: Jike Song <jike.song@intel.com>
Reviewed-by: James Morris <james.l.morris@oracle.com>
Reviewed-by: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
As part of the mdev support, type1 now gets a task reference per
vfio_dma and uses that to get an mm reference for the task while
working on accounting. That's correct, but it's not fast. For some
paths, like vfio_pin_pages_remote(), we know we're only called from
user context, so we can restore the lighter weight calls. In other
cases, we're effectively already testing whether we're in the stored
task context elsewhere, extend this vfio_lock_acct() as well.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Reviewed by: Kirti Wankhede <kwankhede@nvidia.com>
Patch series "mm: unexport __get_user_pages_unlocked()".
This patch series continues the cleanup of get_user_pages*() functions
taking advantage of the fact we can now pass gup_flags as we please.
It firstly adds an additional 'locked' parameter to
get_user_pages_remote() to allow for its callers to utilise
VM_FAULT_RETRY functionality. This is necessary as the invocation of
__get_user_pages_unlocked() in process_vm_rw_single_vec() makes use of
this and no other existing higher level function would allow it to do
so.
Secondly existing callers of __get_user_pages_unlocked() are replaced
with the appropriate higher-level replacement -
get_user_pages_unlocked() if the current task and memory descriptor are
referenced, or get_user_pages_remote() if other task/memory descriptors
are referenced (having acquiring mmap_sem.)
This patch (of 2):
Add a int *locked parameter to get_user_pages_remote() to allow
VM_FAULT_RETRY faulting behaviour similar to get_user_pages_[un]locked().
Taking into account the previous adjustments to get_user_pages*()
functions allowing for the passing of gup_flags, we are now in a
position where __get_user_pages_unlocked() need only be exported for his
ability to allow VM_FAULT_RETRY behaviour, this adjustment allows us to
subsequently unexport __get_user_pages_unlocked() as well as allowing
for future flexibility in the use of get_user_pages_remote().
[sfr@canb.auug.org.au: merge fix for get_user_pages_remote API change]
Link: http://lkml.kernel.org/r/20161122210511.024ec341@canb.auug.org.au
Link: http://lkml.kernel.org/r/20161027095141.2569-2-lstoakes@gmail.com
Signed-off-by: Lorenzo Stoakes <lstoakes@gmail.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Hugh Dickins <hughd@google.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Mel Gorman <mgorman@techsingularity.net>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Radim Krcmar <rkrcmar@redhat.com>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Passing zero for the size to vfio_find_dma() isn't compatible with
matching the start address of an existing vfio_dma. Doing so triggers a
corner case. In vfio_find_dma(), when the start address is equal to
dma->iova and size is 0, check for the end of search range makes it to
take wrong side of RB-tree. That fails the search even though the address
is present in mapped dma ranges.
In functions pin_pages and unpin_pages, the iova which is being searched
is base address of page to be pinned or unpinned. So here size should be
set to PAGE_SIZE, as argument to vfio_find_dma().
Signed-off-by: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Neo Jia <cjia@nvidia.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Passing zero for the size to vfio_find_dma() isn't compatible with
matching the start address of an existing vfio_dma. Doing so triggers a
corner case. In vfio_find_dma(), when the start address is equal to
dma->iova and size is 0, check for the end of search range makes it to
take wrong side of RB-tree. That fails the search even though the address
is present in mapped dma ranges. Due to this, in vfio_dma_do_unmap(),
while checking boundary conditions, size should be set to 1 for verifying
start address of unmap range.
vfio_find_dma() is also used to verify last address in unmap range with
size = 0, but in that case address to be searched is calculated with
start + size - 1 and so it works correctly.
Signed-off-by: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Neo Jia <cjia@nvidia.com>
[aw: changelog tweak]
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
mdev vendor driver should unregister the iommu notifier since the vfio
iommu can persist beyond the attachment of the mdev group. WARN_ON will
show warning if vendor driver doesn't unregister the notifier and is
forced to follow the implementations steps.
Signed-off-by: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Neo Jia <cjia@nvidia.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Currently vfio_register_notifier assumes that there is only one
notifier chain, which is in vfio_iommu. However, the user might
also be interested in events other than vfio_iommu, for example,
vfio_group. Refactor vfio_{un}register_notifier implementation
to make it feasible.
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Xiao Guangrong <guangrong.xiao@linux.intel.com>
Reviewed-by: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Jike Song <jike.song@intel.com>
[aw: merge with commit 816ca69ea9c7 ("vfio: Fix handling of error returned by 'vfio_group_get_from_dev()'"), remove typedef]
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Added blocking notifier to IOMMU TYPE1 driver to notify vendor drivers
about DMA_UNMAP.
Exported two APIs vfio_register_notifier() and vfio_unregister_notifier().
Notifier should be registered, if external user wants to use
vfio_pin_pages()/vfio_unpin_pages() APIs to pin/unpin pages.
Vendor driver should use VFIO_IOMMU_NOTIFY_DMA_UNMAP action to invalidate
mappings.
Signed-off-by: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Neo Jia <cjia@nvidia.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
VFIO IOMMU drivers are designed for the devices which are IOMMU capable.
Mediated device only uses IOMMU APIs, the underlying hardware can be
managed by an IOMMU domain.
Aim of this change is:
- To use most of the code of TYPE1 IOMMU driver for mediated devices
- To support direct assigned device and mediated device in single module
This change adds pin and unpin support for mediated device to TYPE1 IOMMU
backend module. More details:
- Domain for external user is tracked separately in vfio_iommu structure.
It is allocated when group for first mdev device is attached.
- Pages pinned for external domain are tracked in each vfio_dma structure
for that iova range.
- Page tracking rb-tree in vfio_dma keeps <iova, pfn, ref_count>. Key of
rb-tree is iova, but it actually aims to track pfns.
- On external pin request for an iova, page is pinned once, if iova is
already pinned and tracked, ref_count is incremented.
- External unpin request unpins pages only when ref_count is 0.
- Pinned pages list is used to find pfn from iova and then unpin it.
WARN_ON is added if there are entires in pfn_list while detaching the
group and releasing the domain.
- Page accounting is updated to account in its address space where the
pages are pinned/unpinned, i.e dma->task
- Accouting for mdev device is only done if there is no iommu capable
domain in the container. When there is a direct device assigned to the
container and that domain is iommu capable, all pages are already pinned
during DMA_MAP.
- Page accouting is updated on hot plug and unplug mdev device and pass
through device.
Tested by assigning below combinations of devices to a single VM:
- GPU pass through only
- vGPU device only
- One GPU pass through and one vGPU device
- Linux VM hot plug and unplug vGPU device while GPU pass through device
exist
- Linux VM hot plug and unplug GPU pass through device while vGPU device
exist
Signed-off-by: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Neo Jia <cjia@nvidia.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Add task structure to vfio_dma structure. Task structure is used for:
- During DMA_UNMAP, same task who mapped it or other task who shares same
address space is allowed to unmap, otherwise unmap fails.
QEMU maps few iova ranges initially, then fork threads and from the child
thread calls DMA_UNMAP on previously mapped iova. Since child shares same
address space, DMA_UNMAP is successful.
- Avoid accessing struct mm while process is exiting by acquiring
reference of task's mm during page accounting.
- It is also used to get task mlock capability and rlimit for mlock.
Signed-off-by: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Neo Jia <cjia@nvidia.com>
Reviewed-by: Dong Jia Shi <bjsdjshi@linux.vnet.ibm.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Update arguments of vaddr_get_pfn() to take struct mm_struct *mm as input
argument.
Signed-off-by: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Neo Jia <cjia@nvidia.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Added APIs for pining and unpining set of pages. These call back into
backend iommu module to actually pin and unpin pages.
Added two new callback functions to struct vfio_iommu_driver_ops. Backend
IOMMU module that supports pining and unpinning pages for mdev devices
should provide these functions.
Renamed static functions in vfio_type1_iommu.c to resolve conflicts
Signed-off-by: Kirti Wankhede <kwankhede@nvidia.com>
Signed-off-by: Neo Jia <cjia@nvidia.com>
Reviewed-by: Dong Jia Shi <bjsdjshi@linux.vnet.ibm.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
This function cannot actually be called with npage = 0, so in practice
this doesn't return an uninitialized value.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Many IOMMUs support multiple page table formats, meaning that any given
domain may only support a subset of the hardware page sizes presented in
iommu_ops->pgsize_bitmap. There are also certain use-cases where the
creator of a domain may want to control which page sizes are used, for
example to force the use of hugepage mappings to reduce pagetable walk
depth.
To this end, add a per-domain pgsize_bitmap to represent the subset of
page sizes actually in use, to make it possible for domains with
different requirements to coexist.
Signed-off-by: Will Deacon <will.deacon@arm.com>
[rm: hijacked and rebased original patch with new commit message]
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Calling return copy_to_user(...) in an ioctl will not
do the right thing if there's a pagefault:
copy_to_user returns the number of bytes not copied
in this case.
Fix up vfio to do
return copy_to_user(...)) ?
-EFAULT : 0;
everywhere.
Cc: stable@vger.kernel.org
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
The flags entry is there to tell the user that some
optional information is available.
Since we report the iova_pgsizes signal it to the user
by setting the flags to VFIO_IOMMU_INFO_PGSIZES.
Signed-off-by: Pierre Morel <pmorel@linux.vnet.ibm.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Current vfio_pgsize_bitmap code hides the supported IOMMU page
sizes smaller than PAGE_SIZE. As a result, in case the IOMMU
does not support PAGE_SIZE page, the alignment check on map/unmap
is done with larger page sizes, if any. This can fail although
mapping could be done with pages smaller than PAGE_SIZE.
This patch modifies vfio_pgsize_bitmap implementation so that,
in case the IOMMU supports page sizes smaller than PAGE_SIZE
we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
That way the user will be able to map/unmap buffers whose size/
start address is aligned with PAGE_SIZE. Pinning code uses that
granularity while iommu driver can use the sub-PAGE_SIZE size
to map the buffer.
Signed-off-by: Eric Auger <eric.auger@linaro.org>
Acked-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
IOMMU operations can be expensive and it's not very difficult for a
user to give us a lot of work to do for a map or unmap operation.
Killing a large VM will vfio assigned devices can result in soft
lockups and IOMMU tracing shows that we can easily spend 80% of our
time with need-resched set. A sprinkling of conf_resched() calls
after map and unmap calls has a very tiny affect on performance
while resulting in traces with <1% of calls overflowing into needs-
resched.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
We currently map invalid and reserved pages, such as often occur from
mapping MMIO regions of a VM through the IOMMU, using single pages.
There's really no reason we can't instead follow the methodology we
use for normal pages and find the largest possible physically
contiguous chunk for mapping. The only difference is that we don't
do locked memory accounting for these since they're not back by RAM.
In most applications this will be a very minor improvement, but when
graphics and GPGPU devices are in play, MMIO BARs become non-trivial.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
When unmapping DMA entries we try to rely on the IOMMU API behavior
that allows the IOMMU to unmap a larger area than requested, up to
the size of the original mapping. This works great when the IOMMU
supports superpages *and* they're in use. Otherwise, each PAGE_SIZE
increment is unmapped separately, resulting in poor performance.
Instead we can use the IOVA-to-physical-address translation provided
by the IOMMU API and unmap using the largest contiguous physical
memory chunk available, which is also how vfio/type1 would have
mapped the region. For a synthetic 1TB guest VM mapping and shutdown
test on Intel VT-d (2M IOMMU pagesize support), this achieves about
a 30% overall improvement mapping standard 4K pages, regardless of
IOMMU superpage enabling, and about a 40% improvement mapping 2M
hugetlbfs pages when IOMMU superpages are not available. Hugetlbfs
with IOMMU superpages enabled is effectively unchanged.
Unfortunately the same algorithm does not work well on IOMMUs with
fine-grained superpages, like AMD-Vi, costing about 25% extra since
the IOMMU will automatically unmap any power-of-two contiguous
mapping we've provided it. We add a routine and a domain flag to
detect this feature, leaving AMD-Vi unaffected by this unmap
optimization.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
This pull-request includes:
* Change in the IOMMU-API to convert the former iommu_domain_capable
function to just iommu_capable
* Various fixes in handling RMRR ranges for the VT-d driver (one fix
requires a device driver core change which was acked
by Greg KH)
* The AMD IOMMU driver now assigns and deassigns complete alias groups
to fix issues with devices using the wrong PCI request-id
* MMU-401 support for the ARM SMMU driver
* Multi-master IOMMU group support for the ARM SMMU driver
* Various other small fixes all over the place
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2.0.22 (GNU/Linux)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=OMOt
-----END PGP SIGNATURE-----
Merge tag 'iommu-updates-v3.18' of git://git.kernel.org/pub/scm/linux/kernel/git/joro/iommu
Pull IOMMU updates from Joerg Roedel:
"This pull-request includes:
- change in the IOMMU-API to convert the former iommu_domain_capable
function to just iommu_capable
- various fixes in handling RMRR ranges for the VT-d driver (one fix
requires a device driver core change which was acked by Greg KH)
- the AMD IOMMU driver now assigns and deassigns complete alias
groups to fix issues with devices using the wrong PCI request-id
- MMU-401 support for the ARM SMMU driver
- multi-master IOMMU group support for the ARM SMMU driver
- various other small fixes all over the place"
* tag 'iommu-updates-v3.18' of git://git.kernel.org/pub/scm/linux/kernel/git/joro/iommu: (41 commits)
iommu/vt-d: Work around broken RMRR firmware entries
iommu/vt-d: Store bus information in RMRR PCI device path
iommu/vt-d: Only remove domain when device is removed
driver core: Add BUS_NOTIFY_REMOVED_DEVICE event
iommu/amd: Fix devid mapping for ivrs_ioapic override
iommu/irq_remapping: Fix the regression of hpet irq remapping
iommu: Fix bus notifier breakage
iommu/amd: Split init_iommu_group() from iommu_init_device()
iommu: Rework iommu_group_get_for_pci_dev()
iommu: Make of_device_id array const
amd_iommu: do not dereference a NULL pointer address.
iommu/omap: Remove omap_iommu unused owner field
iommu: Remove iommu_domain_has_cap() API function
IB/usnic: Convert to use new iommu_capable() API function
vfio: Convert to use new iommu_capable() API function
kvm: iommu: Convert to use new iommu_capable() API function
iommu/tegra: Convert to iommu_capable() API function
iommu/msm: Convert to iommu_capable() API function
iommu/vt-d: Convert to iommu_capable() API function
iommu/fsl: Convert to iommu_capable() API function
...
VFIO allows devices to be safely handed off to userspace by putting
them behind an IOMMU configured to ensure DMA and interrupt isolation.
This enables userspace KVM clients, such as kvmtool and qemu, to further
map the device into a virtual machine.
With IOMMUs such as the ARM SMMU, it is then possible to provide SMMU
translation services to the guest operating system, which are nested
with the existing translation installed by VFIO. However, enabling this
feature means that the IOMMU driver must be informed that the VFIO domain
is being created for the purposes of nested translation.
This patch adds a new IOMMU type (VFIO_TYPE1_NESTING_IOMMU) to the VFIO
type-1 driver. The new IOMMU type acts identically to the
VFIO_TYPE1v2_IOMMU type, but additionally sets the DOMAIN_ATTR_NESTING
attribute on its IOMMU domains.
Cc: Joerg Roedel <joro@8bytes.org>
Signed-off-by: Will Deacon <will.deacon@arm.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Coverity reports use of a tained scalar used as a loop boundary.
For the most part, any values passed from userspace for a DMA mapping
size, IOVA, or virtual address are valid, with some alignment
constraints. The size is ultimately bound by how many pages the user
is able to lock, IOVA is tested by the IOMMU driver when doing a map,
and the virtual address needs to pass get_user_pages. The only
problem I can find is that we do expect the __u64 user values to fit
within our variables, which might not happen on 32bit platforms. Add
a test for this and return error on overflow. Also propagate use of
the type-correct local variables throughout the function.
The above also points to the 'end' variable, which can be zero if
we're operating at the very top of the address space. We try to
account for this, but our loop botches it. Rework the loop to use
the remaining size as our loop condition rather than the IOVA vs end.
Detected by Coverity: CID 714659
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
- Allow the vfio-type1 IOMMU to support multiple domains within a container
- Plumb path to query whether all domains are cache-coherent
- Wire query into kvm-vfio device to avoid KVM x86 WBINVD emulation
- Always select CONFIG_ANON_INODES, vfio depends on it (Arnd)
The first patch also makes the vfio-type1 IOMMU driver completely independent
of the bus_type of the devices it's handling, which enables it to be used for
both vfio-pci and a future vfio-platform (and hopefully combinations involving
both simultaneously).
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1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=mnzO
-----END PGP SIGNATURE-----
Merge tag 'vfio-v3.15-rc1' of git://github.com/awilliam/linux-vfio
Pull VFIO updates from Alex Williamson:
"VFIO updates for v3.15 include:
- Allow the vfio-type1 IOMMU to support multiple domains within a
container
- Plumb path to query whether all domains are cache-coherent
- Wire query into kvm-vfio device to avoid KVM x86 WBINVD emulation
- Always select CONFIG_ANON_INODES, vfio depends on it (Arnd)
The first patch also makes the vfio-type1 IOMMU driver completely
independent of the bus_type of the devices it's handling, which
enables it to be used for both vfio-pci and a future vfio-platform
(and hopefully combinations involving both simultaneously)"
* tag 'vfio-v3.15-rc1' of git://github.com/awilliam/linux-vfio:
vfio: always select ANON_INODES
kvm/vfio: Support for DMA coherent IOMMUs
vfio: Add external user check extension interface
vfio/type1: Add extension to test DMA cache coherence of IOMMU
vfio/iommu_type1: Multi-IOMMU domain support
Commit bf6bddf192 ("mm: introduce compaction and migration for
ballooned pages") introduces page_count(page) into memory compaction
which dereferences page->first_page if PageTail(page).
This results in a very rare NULL pointer dereference on the
aforementioned page_count(page). Indeed, anything that does
compound_head(), including page_count() is susceptible to racing with
prep_compound_page() and seeing a NULL or dangling page->first_page
pointer.
This patch uses Andrea's implementation of compound_trans_head() that
deals with such a race and makes it the default compound_head()
implementation. This includes a read memory barrier that ensures that
if PageTail(head) is true that we return a head page that is neither
NULL nor dangling. The patch then adds a store memory barrier to
prep_compound_page() to ensure page->first_page is set.
This is the safest way to ensure we see the head page that we are
expecting, PageTail(page) is already in the unlikely() path and the
memory barriers are unfortunately required.
Hugetlbfs is the exception, we don't enforce a store memory barrier
during init since no race is possible.
Signed-off-by: David Rientjes <rientjes@google.com>
Cc: Holger Kiehl <Holger.Kiehl@dwd.de>
Cc: Christoph Lameter <cl@linux.com>
Cc: Rafael Aquini <aquini@redhat.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Now that the type1 IOMMU backend can support IOMMU_CACHE, we need to
be able to test whether coherency is currently enforced. Add an
extension for this.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
We currently have a problem that we cannot support advanced features
of an IOMMU domain (ex. IOMMU_CACHE), because we have no guarantee
that those features will be supported by all of the hardware units
involved with the domain over its lifetime. For instance, the Intel
VT-d architecture does not require that all DRHDs support snoop
control. If we create a domain based on a device behind a DRHD that
does support snoop control and enable SNP support via the IOMMU_CACHE
mapping option, we cannot then add a device behind a DRHD which does
not support snoop control or we'll get reserved bit faults from the
SNP bit in the pagetables. To add to the complexity, we can't know
the properties of a domain until a device is attached.
We could pass this problem off to userspace and require that a
separate vfio container be used, but we don't know how to handle page
accounting in that case. How do we know that a page pinned in one
container is the same page as a different container and avoid double
billing the user for the page.
The solution is therefore to support multiple IOMMU domains per
container. In the majority of cases, only one domain will be required
since hardware is typically consistent within a system. However, this
provides us the ability to validate compatibility of domains and
support mixed environments where page table flags can be different
between domains.
To do this, our DMA tracking needs to change. We currently try to
coalesce user mappings into as few tracking entries as possible. The
problem then becomes that we lose granularity of user mappings. We've
never guaranteed that a user is able to unmap at a finer granularity
than the original mapping, but we must honor the granularity of the
original mapping. This coalescing code is therefore removed, allowing
only unmaps covering complete maps. The change in accounting is
fairly small here, a typical QEMU VM will start out with roughly a
dozen entries, so it's arguable if this coalescing was ever needed.
We also move IOMMU domain creation to the point where a group is
attached to the container. An interesting side-effect of this is that
we now have access to the device at the time of domain creation and
can probe the devices within the group to determine the bus_type.
This finally makes vfio_iommu_type1 completely device/bus agnostic.
In fact, each IOMMU domain can host devices on different buses managed
by different physical IOMMUs, and present a single DMA mapping
interface to the user. When a new domain is created, mappings are
replayed to bring the IOMMU pagetables up to the state of the current
container. And of course, DMA mapping and unmapping automatically
traverse all of the configured IOMMU domains.
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Cc: Varun Sethi <Varun.Sethi@freescale.com>
In vfio_iommu_type1.c there is a bug in vfio_dma_do_map, when checking
that pages are not already mapped. Since the check is being done in a
for loop nested within the main loop, breaking out of it does not create
the intended behavior. If the underlying IOMMU driver returns a non-NULL
value, this will be ignored and mapping the DMA range will be attempted
anyway, leading to unpredictable behavior.
This interracts badly with the ARM SMMU driver issue fixed in the patch
that was submitted with the title:
"[PATCH 2/2] ARM: SMMU: return NULL on error in arm_smmu_iova_to_phys"
Both fixes are required in order to use the vfio_iommu_type1 driver
with an ARM SMMU.
This patch refactors the function slightly, in order to also make this
kind of bug less likely.
Signed-off-by: Antonios Motakis <a.motakis@virtualopensystems.com>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Alex Williamson