This patch fixes a potential small window that the DMA page table might
be incomplete or invalid when the guest sends domain/context
invalidations to a device. This can cause random DMA errors for
assigned devices.
This is a major change to the VT-d shadow page walking logic. It
includes but is not limited to:
- For each VTDAddressSpace, now we maintain what IOVA ranges we have
mapped and what we have not. With that information, now we only send
MAP or UNMAP when necessary. Say, we don't send MAP notifies if we
know we have already mapped the range, meanwhile we don't send UNMAP
notifies if we know we never mapped the range at all.
- Introduce vtd_sync_shadow_page_table[_range] APIs so that we can call
in any places to resync the shadow page table for a device.
- When we receive domain/context invalidation, we should not really run
the replay logic, instead we use the new sync shadow page table API to
resync the whole shadow page table without unmapping the whole
region. After this change, we'll only do the page walk once for each
domain invalidations (before this, it can be multiple, depending on
number of notifiers per address space).
While at it, the page walking logic is also refactored to be simpler.
CC: QEMU Stable <qemu-stable@nongnu.org>
Reported-by: Jintack Lim <jintack@cs.columbia.edu>
Tested-by: Jintack Lim <jintack@cs.columbia.edu>
Signed-off-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Introduce a simplest iova tree implementation based on GTree.
CC: QEMU Stable <qemu-stable@nongnu.org>
Signed-off-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
This patch only modifies the trace points.
Previously we were tracing page walk levels. They are redundant since
we have page mask (size) already. Now we trace something much more
useful which is the domain ID of the page walking. That can be very
useful when we trace more than one devices on the same system, so that
we can know which map is for which domain.
CC: QEMU Stable <qemu-stable@nongnu.org>
Signed-off-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
We pass in the VTDAddressSpace too. It'll be used in the follow up
patches.
CC: QEMU Stable <qemu-stable@nongnu.org>
Signed-off-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
During the recursive page walking of IOVA page tables, some stack
variables are constant variables and never changed during the whole page
walking procedure. Isolate them into a struct so that we don't need to
pass those contants down the stack every time and multiple times.
CC: QEMU Stable <qemu-stable@nongnu.org>
Signed-off-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
For UNMAP-only IOMMU notifiers, we don't need to walk the page tables.
Fasten that procedure by skipping the page table walk. That should
boost performance for UNMAP-only notifiers like vhost.
CC: QEMU Stable <qemu-stable@nongnu.org>
Signed-off-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
SECURITY IMPLICATION: this patch fixes a potential race when multiple
threads access the IOMMU IOTLB cache.
Add a per-iommu big lock to protect IOMMU status. Currently the only
thing to be protected is the IOTLB/context cache, since that can be
accessed even without BQL, e.g., in IO dataplane.
Note that we don't need to protect device page tables since that's fully
controlled by the guest kernel. However there is still possibility that
malicious drivers will program the device to not obey the rule. In that
case QEMU can't really do anything useful, instead the guest itself will
be responsible for all uncertainties.
CC: QEMU Stable <qemu-stable@nongnu.org>
Reported-by: Fam Zheng <famz@redhat.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
That is not really necessary. Removing that node struct and put the
list entry directly into VTDAddressSpace. It simplfies the code a lot.
Since at it, rename the old notifiers_list into vtd_as_with_notifiers.
CC: QEMU Stable <qemu-stable@nongnu.org>
Signed-off-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
SECURITY IMPLICATION: without this patch, any guest with both assigned
device and a vIOMMU might encounter stale IO page mappings even if guest
has already unmapped the page, which may lead to guest memory
corruption. The stale mappings will only be limited to the guest's own
memory range, so it should not affect the host memory or other guests on
the host.
During IOVA page table walking, there is a special case when the PSI
covers one whole PDE (Page Directory Entry, which contains 512 Page
Table Entries) or more. In the past, we skip that entry and we don't
notify the IOMMU notifiers. This is not correct. We should send UNMAP
notification to registered UNMAP notifiers in this case.
For UNMAP only notifiers, this might cause IOTLBs cached in the devices
even if they were already invalid. For MAP/UNMAP notifiers like
vfio-pci, this will cause stale page mappings.
This special case doesn't trigger often, but it is very easy to be
triggered by nested device assignments, since in that case we'll
possibly map the whole L2 guest RAM region into the device's IOVA
address space (several GBs at least), which is far bigger than normal
kernel driver usages of the device (tens of MBs normally).
Without this patch applied to L1 QEMU, nested device assignment to L2
guests will dump some errors like:
qemu-system-x86_64: VFIO_MAP_DMA: -17
qemu-system-x86_64: vfio_dma_map(0x557305420c30, 0xad000, 0x1000,
0x7f89a920d000) = -17 (File exists)
CC: QEMU Stable <qemu-stable@nongnu.org>
Acked-by: Jason Wang <jasowang@redhat.com>
[peterx: rewrite the commit message]
Signed-off-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Ross Zwisler <ross.zwisler@linux.intel.com>
Fixes: commit da6789c27c ("nvdimm: add a macro for property "label-size"")
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Igor Mammedov <imammedo@redhat.com>
Cc: Haozhong Zhang <haozhong.zhang@intel.com>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
This patch reports the protocol feature that is only advertised by
QEMU if the device implements the config ops.
Signed-off-by: Changpeng Liu <changpeng.liu@intel.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
The warning is
hw/virtio/vhost-user.c:1319:26: error: suggest braces
around initialization of subobject [-Werror,-Wmissing-braces]
VhostUserMsg msg = { 0 };
^
{}
While the original code is correct, and technically exactly correct
as per ISO C89, both GCC and Clang support plain empty set of braces
as an extension.
Cc: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
The response to a VHOST_USER_POSTCOPY_ADVISE contains a fd but doesn't
actually contain any data. FIx vu_message_write so that it doesn't
do a 0-byte write() call, since this was ending up with rc=0
that was confusing the error handling code.
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Use qemu_set_nonblock rather than a simple fcntl; cleaner
and I have no reason to change other flags.
Reported-by: Peter Maydell <peter.maydell@linaro.org>
Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
This patch introduces the support for setting memory region
based host notifiers for virtio device. This is helpful when
using a hardware accelerator for a virtio device, because
hardware heavily depends on the notification, this will allow
the guest driver in the VM to notify the hardware directly.
Signed-off-by: Tiwei Bie <tiwei.bie@intel.com>
Reviewed-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
qmp_to_opts() used to be a method of QMPTestCase, but recently we
started to add more Python test cases that don't make use of
QMPTestCase. In order to make the method usable there, move it to VM.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
This adds a minimal query-jobs implementation that shouldn't pose many
design questions. It can later be extended to expose more information,
and especially job-specific information.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This adds a QMP event that is emitted whenever a job transitions from
one status to another.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
This adds a separate schema file for all job-related definitions that
aren't tied to the block layer.
For a start, move the enums JobType, JobStatus and JobVerb.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
BlockJob has fields .offset and .len, which are actually misnomers today
because they are no longer tied to block device sizes, but just progress
counters. As such they make a lot of sense in generic Jobs.
This patch moves the fields to Job and renames them to .progress_current
and .progress_total to describe their function better.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
The transition to the READY state was still performed in the BlockJob
layer, in the same function that sent the BLOCK_JOB_READY QMP event.
This patch brings the state transition to the Job layer and implements
the QMP event using a notifier called from the Job layer, like we
already do for other events related to state transitions.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Instead of having a 'bool ready' in BlockJob, add a function that
derives its value from the job status.
At the same time, this fixes the behaviour to match what the QAPI
documentation promises for query-block-job: 'true if the job may be
completed'. When the ready flag was introduced in commit ef6dbf1e46,
the flag never had to be reset to match the description because after
being ready, the jobs would immediately complete and disappear.
Job transactions and manual job finalisation were introduced only later.
With these changes, jobs may stay around even after having completed
(and they are not ready to be completed a second time), however their
patches forgot to reset the ready flag.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Now that we cancel all jobs and not only block jobs on shutdown, doing
that in bdrv_close_all() isn't really appropriate any more. Move the
job_cancel_sync_all() call to the callers, and only assert that there
are no job running in bdrv_close_all().
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
This moves the logic that implements job transactions from BlockJob to
Job.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
This doesn't actually move any transaction code to Job yet, but it
renames the type for transactions from BlockJobTxn to JobTxn and makes
them contain Jobs rather than BlockJobs
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
block_job_finish_sync() doesn't contain anything block job specific any
more, so it can be moved to Job.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
This moves the .complete callback that tells a READY job to complete
from BlockJobDriver to JobDriver. The wrapper function job_complete()
doesn't require anything block job specific any more and can be moved
to Job.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
block_job_drain() contains a blk_drain() call which cannot be moved to
Job, so add a new JobDriver callback JobDriver.drain which has a common
implementation for all BlockJobs. In addition to this we keep the
existing BlockJobDriver.drain callback that is called by the common
drain implementation for all block jobs.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
block_job_cancel_async() did two things that were still block job
specific:
* Setting job->force. This field makes sense on the Job level, so we can
just move it. While at it, rename it to job->force_cancel to make its
purpose more obvious.
* Resetting the I/O status. This can't be moved because generic Jobs
don't have an I/O status. What the function really implements is a
user resume, except without entering the coroutine. Consequently, it
makes sense to call the .user_resume driver callback here which
already resets the I/O status.
The old block_job_cancel_async() has two separate if statements that
check job->iostatus != BLOCK_DEVICE_IO_STATUS_OK and job->user_paused.
However, the former condition always implies the latter (as is
asserted in block_job_iostatus_reset()), so changing the explicit call
of block_job_iostatus_reset() on the former condition with the
.user_resume callback on the latter condition is equivalent and
doesn't need to access any BlockJob specific state.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
This moves the finalisation of a single job from BlockJob to Job.
Some part of this code depends on job transactions, and job transactions
call this code, we introduce some temporary calls from Job functions to
BlockJob ones. This will be fixed once transactions move to Job, too.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Go through the Job layer in order to send QMP events. For the moment,
these functions only call a notifier in the BlockJob layer that sends
the existing commands.
This uses notifiers rather than JobDriver callbacks because internal
users of jobs won't receive QMP events, but might still be interested
in getting notified for the events.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
block_job_event_pending() doesn't only send a QMP event, but it also
transitions to the PENDING state. Split the function so that we get one
part only sending the event (like other block_job_event_* functions) and
another part that does the state transition.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
This renames the BlockJobCreateFlags constants, moves a few JOB_INTERNAL
checks to job_create() and the auto_{finalize,dismiss} fields from
BlockJob to Job.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Since we introduced an explicit status to block job, BlockJob.completed
is redundant because it can be derived from the status. Remove the field
from BlockJob and add a function to derive it from the status at the Job
level.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
While we already moved the state related to job pausing to Job, the
functions to do were still BlockJob only. This commit moves them over to
Job.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
There is nothing block layer specific about block_job_sleep_ns(), so
move the function to Job.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
This commit moves some core functions for dealing with the job coroutine
from BlockJob to Job. This includes primarily entering the coroutine
(both for the first and reentering) and yielding explicitly and at pause
points.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
Move the defer_to_main_loop functionality from BlockJob to Job.
The code can be simplified because we can use job->aio_context in
job_defer_to_main_loop_bh() now, instead of having to access the
BlockDriverState.
Probably taking the data->aio_context lock in addition was already
unnecessary in the old code because we didn't actually make use of
anything protected by the old AioContext except getting the new
AioContext, in case it changed between scheduling the BH and running it.
But it's certainly unnecessary now that the BDS isn't accessed at all
any more.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
When block jobs need an AioContext, they just take it from their main
block node. Generic jobs don't have a main block node, so we need to
assign them an AioContext explicitly.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
We cannot yet move the whole logic around job cancelling to Job because
it depends on quite a few other things that are still only in BlockJob,
but we can move the cancelled field at least.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
This moves reference counting from BlockJob to Job.
In order to keep calling the BlockJob cleanup code when the job is
deleted via job_unref(), introduce a new JobDriver.free callback. Every
block job must use block_job_free() for this callback, this is asserted
in block_job_create().
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
This moves BlockJob.status and the closely related functions
(block_)job_state_transition() and (block_)job_apply_verb to Job. The
two QAPI enums are renamed to JobStatus and JobVerb.
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
Reviewed-by: Eric Blake <eblake@redhat.com>
This moves the job list from BlockJob to Job. Now we can check for
duplicate IDs in job_create().
Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Reviewed-by: John Snow <jsnow@redhat.com>
