Improve the source side flow upon pre_copy as of below.
- Prepare the stop_copy buffers as part of moving to pre_copy.
- Send to the target a record that includes the expected
stop_copy size to let it optimize its stop_copy flow as well.
As for sending the target this new record type (i.e.
MLX5_MIGF_HEADER_TAG_STOP_COPY_SIZE) we split the current 64 header
flags bits into 32 flags bits and another 32 tag bits, each record may
have a tag and a flag whether it's optional or mandatory. Optional
records will be ignored in the target.
The above reduces the downtime upon stop_copy as the relevant data stuff
is prepared ahead as part of pre_copy.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20230124144955.139901-3-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Use GFP_KERNEL_ACCOUNT for userspace persistent allocations.
The GFP_KERNEL_ACCOUNT option lets the memory allocator know that this
is untrusted allocation triggered from userspace and should be a subject
of kmem accounting, and as such it is controlled by the cgroup
mechanism.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Link: https://lore.kernel.org/r/20230108154427.32609-5-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Use GFP_KERNEL_ACCOUNT for userspace persistent allocations.
The GFP_KERNEL_ACCOUNT option lets the memory allocator know that this
is untrusted allocation triggered from userspace and should be a subject
of kmem accounting, and as such it is controlled by the cgroup
mechanism.
The way to find the relevant allocations was for example to look at the
close_device function and trace back all the kfrees to their
allocations.
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Link: https://lore.kernel.org/r/20230108154427.32609-4-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Allow loading of larger images than 512 MB by dropping the arbitrary
hard-coded value that we have today and move to use the max device
loading value which is for now 4GB.
As part of that we move to use the GFP_KERNEL_ACCOUNT option upon
allocating the persistent data of mlx5 and rely on the cgroup to provide
the memory limit for the given user.
The GFP_KERNEL_ACCOUNT option lets the memory allocator know that this
is untrusted allocation triggered from userspace and should be a subject
of kmem accounting, and as such it is controlled by the cgroup
mechanism.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Link: https://lore.kernel.org/r/20230108154427.32609-3-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Prevent calling roundup_pow_of_two() with value of 0 as it causes the
below UBSAN note.
Move this code and its few extra related lines to be called only when
it's really applicable.
UBSAN: shift-out-of-bounds in ./include/linux/log2.h:57:13
shift exponent 64 is too large for 64-bit type 'long unsigned int'
CPU: 15 PID: 1639 Comm: live_migration Not tainted 6.1.0-rc4 #1116
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009),
BIOS rel-1.13.0-0-gf21b5a4aeb02-prebuilt.qemu.org 04/01/2014
Call Trace:
<TASK>
dump_stack_lvl+0x45/0x59
ubsan_epilogue+0x5/0x36
__ubsan_handle_shift_out_of_bounds.cold+0x61/0xef
? lock_is_held_type+0x98/0x110
? rcu_read_lock_sched_held+0x3f/0x70
mlx5vf_create_rc_qp.cold+0xe4/0xf2 [mlx5_vfio_pci]
mlx5vf_start_page_tracker+0x769/0xcd0 [mlx5_vfio_pci]
vfio_device_fops_unl_ioctl+0x63f/0x700 [vfio]
__x64_sys_ioctl+0x433/0x9a0
do_syscall_64+0x3d/0x90
entry_SYSCALL_64_after_hwframe+0x63/0xcd
</TASK>
Fixes: 79c3cf2799 ("vfio/mlx5: Init QP based resources for dirty tracking")
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Link: https://lore.kernel.org/r/20230108154427.32609-2-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Pull VFIO updates from Alex Williamson:
- Replace deprecated git://github.com link in MAINTAINERS (Palmer
Dabbelt)
- Simplify vfio/mlx5 with module_pci_driver() helper (Shang XiaoJing)
- Drop unnecessary buffer from ACPI call (Rafael Mendonca)
- Correct latent missing include issue in iova-bitmap and fix support
for unaligned bitmaps. Follow-up with better fix through refactor
(Joao Martins)
- Rework ccw mdev driver to split private data from parent structure,
better aligning with the mdev lifecycle and allowing us to remove a
temporary workaround (Eric Farman)
- Add an interface to get an estimated migration data size for a
device, allowing userspace to make informed decisions, ex. more
accurately predicting VM downtime (Yishai Hadas)
- Fix minor typo in vfio/mlx5 array declaration (Yishai Hadas)
- Simplify module and Kconfig through consolidating SPAPR/EEH code and
config options and folding virqfd module into main vfio module (Jason
Gunthorpe)
- Fix error path from device_register() across all vfio mdev and sample
drivers (Alex Williamson)
- Define migration pre-copy interface and implement for vfio/mlx5
devices, allowing portions of the device state to be saved while the
device continues operation, towards reducing the stop-copy state size
(Jason Gunthorpe, Yishai Hadas, Shay Drory)
- Implement pre-copy for hisi_acc devices (Shameer Kolothum)
- Fixes to mdpy mdev driver remove path and error path on probe (Shang
XiaoJing)
- vfio/mlx5 fixes for incorrect return after copy_to_user() fault and
incorrect buffer freeing (Dan Carpenter)
* tag 'vfio-v6.2-rc1' of https://github.com/awilliam/linux-vfio: (42 commits)
vfio/mlx5: error pointer dereference in error handling
vfio/mlx5: fix error code in mlx5vf_precopy_ioctl()
samples: vfio-mdev: Fix missing pci_disable_device() in mdpy_fb_probe()
hisi_acc_vfio_pci: Enable PRE_COPY flag
hisi_acc_vfio_pci: Move the dev compatibility tests for early check
hisi_acc_vfio_pci: Introduce support for PRE_COPY state transitions
hisi_acc_vfio_pci: Add support for precopy IOCTL
vfio/mlx5: Enable MIGRATION_PRE_COPY flag
vfio/mlx5: Fallback to STOP_COPY upon specific PRE_COPY error
vfio/mlx5: Introduce multiple loads
vfio/mlx5: Consider temporary end of stream as part of PRE_COPY
vfio/mlx5: Introduce vfio precopy ioctl implementation
vfio/mlx5: Introduce SW headers for migration states
vfio/mlx5: Introduce device transitions of PRE_COPY
vfio/mlx5: Refactor to use queue based data chunks
vfio/mlx5: Refactor migration file state
vfio/mlx5: Refactor MKEY usage
vfio/mlx5: Refactor PD usage
vfio/mlx5: Enforce a single SAVE command at a time
vfio: Extend the device migration protocol with PRE_COPY
...
Before a SAVE command is issued, a QUERY command is issued in order to
know the device data size.
In case PRE_COPY is used, the above commands are issued while the device
is running. Thus, it is possible that between the QUERY and the SAVE
commands the state of the device will be changed significantly and thus
the SAVE will fail.
Currently, if a SAVE command is failing, the driver will fail the
migration. In the above case, don't fail the migration, but don't allow
for new SAVEs to be executed while the device is in a RUNNING state.
Once the device will be moved to STOP_COPY, SAVE can be executed again
and the full device state will be read.
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20221206083438.37807-14-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
In order to support PRE_COPY, mlx5 driver transfers multiple states
(images) of the device. e.g.: the source VF can save and transfer
multiple states, and the target VF will load them by that order.
This patch implements the changes for the target VF to decompose the
header for each state and to write and load multiple states.
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20221206083438.37807-13-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
During PRE_COPY the migration data FD may have a temporary "end of
stream" that is reached when the initial_bytes were read and no other
dirty data exists yet.
For instance, this may indicate that the device is idle and not
currently dirtying any internal state. When read() is done on this
temporary end of stream the kernel driver should return ENOMSG from
read(). Userspace can wait for more data or consider moving to
STOP_COPY.
To not block the user upon read() and let it get ENOMSG we add a new
state named MLX5_MIGF_STATE_PRE_COPY on the migration file.
In addition, we add the MLX5_MIGF_STATE_SAVE_LAST state to block the
read() once we call the last SAVE upon moving to STOP_COPY.
Any further error will be marked with MLX5_MIGF_STATE_ERROR and the user
won't be blocked.
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20221206083438.37807-12-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
vfio precopy ioctl returns an estimation of data available for
transferring from the device.
Whenever a user is using VFIO_MIG_GET_PRECOPY_INFO, track the current
state of the device, and if needed, append the dirty data to the
transfer FD data. This is done by saving a middle state.
As mlx5 runs the SAVE command asynchronously, make sure to query for
incremental data only once there is no active save command.
Running both in parallel, might end-up with a failure in the incremental
query command on un-tracked vhca.
Also, a middle state will be saved only after the previous state has
finished its SAVE command and has been fully transferred, this prevents
endless use resources.
Co-developed-by: Shay Drory <shayd@nvidia.com>
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20221206083438.37807-11-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
As mentioned in the previous patches, mlx5 is transferring multiple
states when the PRE_COPY protocol is used. This states mechanism
requires the target VM to know the states' size in order to execute
multiple loads. Therefore, add SW header, with the needed information,
for each saved state the source VM is transferring to the target VM.
This patch implements the source VM handling of the headers, following
patch will implement the target VM handling of the headers.
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20221206083438.37807-10-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
In order to support PRE_COPY, mlx5 driver is transferring multiple
states (images) of the device. e.g.: the source VF can save and transfer
multiple states, and the target VF will load them by that order.
The device is saving three kinds of states:
1) Initial state - when the device moves to PRE_COPY state.
2) Middle state - during PRE_COPY phase via VFIO_MIG_GET_PRECOPY_INFO.
There can be multiple states of this type.
3) Final state - when the device moves to STOP_COPY state.
After moving to PRE_COPY state, user is holding the saving migf FD and
can use it. For example: user can start transferring data via read()
callback. Also, user can switch from PRE_COPY to STOP_COPY whenever he
sees it fits. This will invoke saving of final state.
This means that mlx5 VFIO device can be switched to STOP_COPY without
transferring any data in PRE_COPY state. Therefore, when the device
moves to STOP_COPY, mlx5 will store the final state on a dedicated queue
entry on the list.
Co-developed-by: Shay Drory <shayd@nvidia.com>
Signed-off-by: Shay Drory <shayd@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20221206083438.37807-9-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Refactor to use queue based data chunks on the migration file.
The SAVE command adds a chunk to the tail of the queue while the read()
API finds the required chunk and returns its data.
In case the queue is empty but the state of the migration file is
MLX5_MIGF_STATE_COMPLETE, read() may not be blocked but will return 0 to
indicate end of file.
This is a step towards maintaining multiple images and their meta data
(i.e. headers) on the migration file as part of next patches from the
series.
Note:
At that point, we still use a single chunk on the migration file but
becomes ready to support multiple.
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20221206083438.37807-8-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
This patch refactors MKEY usage such as its life cycle will be as of the
migration file instead of allocating/destroying it upon each
SAVE/LOAD command.
This is a preparation step towards the PRE_COPY series where multiple
images will be SAVED/LOADED.
We achieve it by having a new struct named mlx5_vhca_data_buffer which
holds the mkey and its related stuff as of sg_append_table,
allocated_length, etc.
The above fields were taken out from the migration file main struct,
into mlx5_vhca_data_buffer dedicated struct with the proper helpers in
place.
For now we have a single mlx5_vhca_data_buffer per migration file.
However, in coming patches we'll have multiple of them to support
multiple images.
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20221206083438.37807-6-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
This patch refactors PD usage such as its life cycle will be as of the
migration file instead of allocating/destroying it upon each SAVE/LOAD
command.
This is a preparation step towards the PRE_COPY series where multiple
images will be SAVED/LOADED and a single PD can be simply reused.
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20221206083438.37807-5-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Enforce a single SAVE command at a time.
As the SAVE command is an asynchronous one, we must enforce running only
a single command at a time.
This will preserve ordering between multiple calls and protect from
races on the migration file data structure.
This is a must for the next patches from the series where as part of
PRE_COPY we may have multiple images to be saved and multiple SAVE
commands may be issued from different flows.
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20221206083438.37807-4-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Resolve conflicts in drivers/vfio/vfio_main.c by using the iommfd version.
The rc fix was done a different way when iommufd patches reworked this
code.
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
This creates the iommufd_device for the physical VFIO drivers. These are
all the drivers that are calling vfio_register_group_dev() and expect the
type1 code to setup a real iommu_domain against their parent struct
device.
The design gives the driver a choice in how it gets connected to iommufd
by providing bind_iommufd/unbind_iommufd/attach_ioas callbacks to
implement as required. The core code provides three default callbacks for
physical mode using a real iommu_domain. This is suitable for drivers
using vfio_register_group_dev()
Link: https://lore.kernel.org/r/6-v4-42cd2eb0e3eb+335a-vfio_iommufd_jgg@nvidia.com
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Reviewed-by: Alex Williamson <alex.williamson@redhat.com>
Tested-by: Alex Williamson <alex.williamson@redhat.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>
Tested-by: Yu He <yu.he@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Add an option to get migration data size by introducing a new migration
feature named VFIO_DEVICE_FEATURE_MIG_DATA_SIZE.
Upon VFIO_DEVICE_FEATURE_GET the estimated data length that will be
required to complete STOP_COPY is returned.
This option may better enable user space to consider before moving to
STOP_COPY whether it can meet the downtime SLA based on the returned
data.
The patch also includes the implementation for mlx5 and hisi for this
new option to make it feature complete for the existing drivers in this
area.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Reviewed-by: Longfang Liu <liulongfang@huawei.com>
Link: https://lore.kernel.org/r/20221106174630.25909-2-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
vfio_pci_dev_set_needs_reset() inspects the open_count of every device
in the set to determine whether a reset is allowed. The current device
always has open_count == 1 within vfio_pci_core_disable(), effectively
disabling the reset logic. This field is also documented as private in
vfio_device, so it should not be used to determine whether other devices
in the set are open.
Checking for vfio_device_set_open_count() > 1 on the device set fixes
both issues.
After commit 2cd8b14aaa ("vfio/pci: Move to the device set
infrastructure"), failure to create a new file for a device would cause
the reset to be skipped due to open_count being decremented after
calling close_device() in the error path.
After commit eadd86f835 ("vfio: Remove calls to
vfio_group_add_container_user()"), releasing a device would always skip
the reset due to an ordering change in vfio_device_fops_release().
Failing to reset the device leaves it in an unknown state, potentially
causing errors when it is accessed later or bound to a different driver.
This issue was observed with a Radeon RX Vega 56 [1002:687f] (rev c3)
assigned to a Windows guest. After shutting down the guest, unbinding
the device from vfio-pci, and binding the device to amdgpu:
[ 548.007102] [drm:psp_hw_start [amdgpu]] *ERROR* PSP create ring failed!
[ 548.027174] [drm:psp_hw_init [amdgpu]] *ERROR* PSP firmware loading failed
[ 548.027242] [drm:amdgpu_device_fw_loading [amdgpu]] *ERROR* hw_init of IP block <psp> failed -22
[ 548.027306] amdgpu 0000:0a:00.0: amdgpu: amdgpu_device_ip_init failed
[ 548.027308] amdgpu 0000:0a:00.0: amdgpu: Fatal error during GPU init
Fixes: 2cd8b14aaa ("vfio/pci: Move to the device set infrastructure")
Fixes: eadd86f835 ("vfio: Remove calls to vfio_group_add_container_user()")
Signed-off-by: Anthony DeRossi <ajderossi@gmail.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Reviewed-by: Kevin Tian <kevin.tian@intel.com>
Link: https://lore.kernel.org/r/20221110014027.28780-4-ajderossi@gmail.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Report dirty pages from tracker.
It includes:
Querying for dirty pages in a given IOVA range, this is done by
modifying the tracker into the reporting state and supplying the
required range.
Using the CQ event completion mechanism to be notified once data is
ready on the CQ/QP to be processed.
Once data is available turn on the corresponding bits in the bit map.
This functionality will be used as part of the 'log_read_and_clear'
driver callback in the next patches.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20220908183448.195262-9-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Init QP based resources for dirty tracking to be used upon start
logging.
It includes:
Creating the host and firmware RC QPs, move each of them to its expected
state based on the device specification, etc.
Creating the relevant resources which are needed by both QPs as of UAR,
PD, etc.
Creating the host receive side resources as of MKEY, CQ, receive WQEs,
etc.
The above resources are cleaned-up upon stop logging.
The tracker object that will be introduced by next patches will use
those resources.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20220908183448.195262-7-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Introduce the DMA logging feature support in the vfio core layer.
It includes the processing of the device start/stop/report DMA logging
UAPIs and calling the relevant driver 'op' to do the work.
Specifically,
Upon start, the core translates the given input ranges into an interval
tree, checks for unexpected overlapping, non aligned ranges and then
pass the translated input to the driver for start tracking the given
ranges.
Upon report, the core translates the given input user space bitmap and
page size into an IOVA kernel bitmap iterator. Then it iterates it and
call the driver to set the corresponding bits for the dirtied pages in a
specific IOVA range.
Upon stop, the driver is called to stop the previous started tracking.
The next patches from the series will introduce the mlx5 driver
implementation for the logging ops.
Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Link: https://lore.kernel.org/r/20220908183448.195262-6-yishaih@nvidia.com
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
