Close the file when either memory allocation or a file read fails in
amdgpu_plugin_dmabuf_restore.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@igalia.com>
Docstring refers to size_t as 8-byte value while in reality this is
architecture dependent. Clarify it.
Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@igalia.com>
In amdgpu_plugin_drm_restore_file(), dmabuf_fd can reach the
drmPrimeFDToHandle() call as -1 if fdstore_get() fails or if
boinfo->handle is -1 while boinfo->is_import is true. Passing
an invalid fd to libdrm causes a hard restore failure instead
of a retry. Check for -1 and set retry_needed so the restore
coordinator can retry later.
When retry_needed is set, some dmabufs entries are still invalid,
so skip record_completed_work() and restore_bo_contents_drm()
to avoid operating on incomplete state.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
Several error paths in amdgpu_plugin_drm_dump_file() leak
resources. The drmPrimeHandleToFD() and open_img_file() failures
use break instead of goto exit, which skips per-iteration cleanup
and lets the function continue dumping metadata with a stale ret
value. Replace break with proper resource cleanup and goto exit.
The vm_info_entries buffer is not freed when the ioctl or
allocate_vm_entries() calls fail after allocating it, and the
buffer from posix_memalign() is never freed after sdma_copy_bo(),
leaking memory on every loop iteration. Add the missing xfree()
calls.
Check the amdgpu_device_initialize() return value because h_dev
is undefined on failure. Move list_handles_entries cleanup into
the exit block because error paths that goto exit skip the
existing xfree() call after the loop. Add a NULL check before
free_e(rd) because rd allocation can fail before the struct is
initialized.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
Initialize gpu_ids, fds_write, and entries to NULL before
allocating them in init_parallel_restore_cmd() and
init_parallel_restore_cmd_by_head(). The caller declares the
struct on the stack without an initializer, so these members
hold undefined values. If a later xzalloc() fails and the
caller invokes free_parallel_restore_cmd(), it will call
xfree() on those undefined pointers.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
Call amdgpu_device_deinitialize() before continuing when
get_gem_handle() fails and before returning when a matching
handle is found. Both paths skip the existing deinitialize
call at the end of the loop body, leaking the device handle.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
open_img_file() can return NULL after a read_fp() or write_fp()
failure without setting errno. The openat() and fdopen() paths
set errno, but the read/write path does not, so callers using
-errno get zero or a stale value. Set errno to EIO on this path
so errno is always valid when the function returns NULL.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
Close the file descriptor when fdopen() fails, as fdopen does not
take ownership of the fd on failure.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
The exit label in amdgpu_plugin_drm_restore_file() returns
immediately when ret < 0 without freeing the dmabufs array or
deinitializing the amdgpu device handle. Move cleanup before the
error check so resources are always released.
Replace xzalloc() with xmalloc() and memset(0xff) for the dmabufs
array because zero is a valid fd number and would cause the cleanup
loop to close fd 0 (stdin) for unpopulated entries. Initializing
to KFD_INVALID_FD (0xffffffff) lets the exit block tell which fds
were actually opened. Guard amdgpu_device_deinitialize() behind a
dev_initialized flag to avoid calling it when initialization failed,
and close dmabuf_fd when drmPrimeFDToHandle() fails. To prevent
double-closing fds that restore_bo_contents_drm() already closed,
only run the close loop when the BO content restore was not reached.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
Free the protobuf message and buffer when returning early from
amdgpu_plugin_dmabuf_restore() on failure to find a dmabuf fd.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
posix_memalign() returns an error code on failure and does not
guarantee setting the pointer to NULL. Check the return value
instead of the pointer. Since posix_memalign() does not set
errno, assign it explicitly so pr_perror() prints the correct
message. Negate the return code instead of hardcoding -ENOMEM
to preserve the actual error (e.g. EINVAL for bad alignment).
In amdgpu_plugin_drm_dump_file(), replace break with cleanup
and goto exit so per-BO resources are freed on failure.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
When the amdgpu plugin is called for restore of a file img,
we don't initially know if it's a kfd or drm file. So we first
try to open a kfd file, and if that fails, we assume it's
a drm file. That failure leads to a lot of error messages like
Error amdgpu_plugin: amdgpu-kfd-141.img: Failed to open for read
For this normal behaviour.
Add an option on open_img_file to suppress this error when
we aren't sure if the file we're trying to open exists.
Signed-off-by: David Francis <David.Francis@amd.com>
open_img_file can return null if, for example, the dump file
isn't present. Tehre were three places where this wasn't checked.
Check it.
Signed-off-by: David Francis <David.Francis@amd.com>
pr_perror already appends a newline; the explicit \n in the
format string violates the lint rule and breaks CI.
Fixes: 555b257c5e77 ("plugin/amdgpu: Catch error for failure to open drm device")
Signed-off-by: Ahmed Elaidy <elaidya225@gmail.com>
kfd_criu_device_bucket.drm_fd is an unsigned int, so the comparison
device_bucket->drm_fd < 0
was always false.
Fix that, and also make the error message a bit more descriptive.
Signed-off-by: David Francis <David.Francis@amd.com>
In container setups, particularly k8s, cgroup may be used to
restrict what GPUs a process has access to. CRIU may then be used
to checkpoint that process while CRIU itself has full access to
all GPUs.
In this case, the kfd CRIU ioctl will return only a subset of the
ioctls visible through sysfs.
Don't write these devices or any links connected to them to the
tology dump files.
Signed-off-by: David Francis <David.Francis@amd.com>
As of Alpine Linux 3.19, musl libc no longer contains separate
fopen64(), fseeko64(), or ftello64() functions. This causes building
CRIU with amdgpu plugin to fail with the following error:
amdgpu_plugin.c: In function 'parallel_restore_bo_contents':
amdgpu_plugin.c:2286:17: error: implicit declaration of function 'fseeko64'; did you mean 'fseeko'? [-Wimplicit-function-declaration]
2286 | fseeko64(bo_contents_fp, entry->read_offset + offset, SEEK_SET);
| ^~~~~~~~
| fseeko
make[2]: *** [Makefile:31: amdgpu_plugin.so] Error 1
make[1]: *** [Makefile:363: amdgpu_plugin] Error 2
To fix this, add the missing $(DEFINES) to plugin builds, and since we
always compile with _FILE_OFFSET_BITS=64, we don't need the 64 suffix.
Fixes: #2826
Suggested-by: Andrei Vagin <avagin@google.com>
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
On AMD Instinct MI300 systems, restoring a large GPU application can
fail because the checkpoint size is too large and the maximum value of
an offset (with integer type) is insufficient. This problem occurs when
the total size of all buffer objects exceeds int max, not because any
single buffer is too large, but it can also happen with a large number
of small buffers.
Fixes: #2812
Signed-off-by: Yanning Yang <yangyanning@sjtu.edu.cn>
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
This patch updates drm.h and amdgpu_drm.h kernel headers,
and adds drm_mode.h (included by drm.h) from the rocm-7.1.0
release tag.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
amdgpu_plugin_drm.c:167:6: error: variable 'num_bos' set but not used [-Werror,-Wunused-but-set-variable]
167 | int num_bos = 0;
|
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
Add a comment that explains the purpose of `retry_needed`.
Co-authored-by: Andrei Vagin <avagin@google.com>
Signed-off-by: David Francis <David.Francis@amd.com>
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
Co-authored-by: Andrei Vagin <avagin@google.com>
Signed-off-by: David Francis <David.Francis@amd.com>
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
Use `return 0` on success in `post_dump_dmabuf_check()` for consistency
with other functions.
Co-authored-by: Andrei Vagin <avagin@google.com>
Signed-off-by: David Francis <David.Francis@amd.com>
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
These pr_info lines begin with "CC3" and "TWI" were not meant to be
included in the patch.
Co-authored-by: Andrei Vagin <avagin@google.com>
Signed-off-by: David Francis <David.Francis@amd.com>
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
amdgpu libraries that use dmabuf fd to share GPU memory between
processes close the dmabuf fds immediately after using them.
However, it is possible that checkpoint of a process catches one
of the dmabuf fds open. In that case, the amdgpu plugin needs
to handle it.
The checkpoint of the dmabuf fd does require the device file
it was exported from to have already been dumped
To identify which device this dmabuf fd was exprted from, attempt
to import it on each device, then record the dmabuf handle
it imports as. This handle can be used to restore it.
Signed-off-by: David Francis <David.Francis@amd.com>
The amdgpu plugin was counting how many files were checkpointed
to determine when it should close the device files.
The number of device files is not consistent; a process may
have multiple copies of the drm device files open.
Instead of doing this counting, add a new callback after all
files are checkpointed, so plugins can clean up their
resources at an appropriate time.
Signed-off-by: David Francis <David.Francis@amd.com>
Buffer objects held by the amdgpu drm driver are checkpointed with
the new BO_INFO and MAPPING_INFO ioctls/ioctl options. Handling
is in amdgpu_plugin_drm.h
Handling of imported buffer objects may require dmabuf fds to be
transferred between processes. These occur over fdstore, with the
handle-fstore id relationships kept in shread memory. There is a
new plugin callback: RESTORE_INIT to create the shared memory.
During checkpoint, track shared buffer objects, so that buffer objects
that are shared across processes can be identified.
During restore, track which buffer objects have been restored. Retry
restore of a drm file if a buffer object is imported and the
original has not been exported yet. Skip buffer objects that have
already been completed or cannot be completed in the current restore.
So drm code can use sdma_copy_bo, that function no longer requires
kfd bo structs
Update the protobuf messages with new amdgpu drm information.
Signed-off-by: David Francis <David.Francis@amd.com>
The amdgpu plugin usually calls drm ioctls through the libdrm
wrappers. However, amdgpu restore requires dealing with dmabufs
and gem handles directly, which means drm ioctls must be
called directly.
Add the drm.h header (from the kernel's uapi).
Signed-off-by: David Francis <David.Francis@amd.com>
For amdgpu plugin to call the new amdgpu drm CRIU ioctls, it needs
the amdgpu drm header file, copied from the kernel's includes.
Signed-off-by: David Francis <David.Francis@amd.com>
amdgpu dmabuf CRIU requires the ability of the amdgpu plugin
to retry.
Change files_ext.c to read a response of 1 from a plugin restore
function to mean retry.
Signed-off-by: David Francis <David.Francis@amd.com>
This patch implements the entire logic to enable the offloading of
buffer object content restoration.
The goal of this patch is to offload the buffer object content
restoration to the main CRIU process so that this restoration can occur
in parallel with other restoration logic (mainly the restoration of
memory state in the restore blob, which is time-consuming) to speed up
the restore phase. The restoration of buffer object content usually
takes a significant amount of time for GPU applications, so
parallelizing it with other operations can reduce the overall restore
time.
It has three parts: the first replaces the restoration of buffer objects
in the target process by sending a parallel restore command to the main
CRIU process; the second implements the POST_FORKING hook in the amdgpu
plugin to enable buffer object content restoration in the main CRIU
process; the third stops the parallel thread in the RESUME_DEVICES_LATE
hook.
This optimization only focuses on the single-process situation (common
case). In other scenarios, it will turn to the original method. This is
achieved with the new `parallel_disabled` flag.
Signed-off-by: Yanning Yang <yangyanning@sjtu.edu.cn>
Currently the restore of buffer object comsumes a significant amount of
time. However, this part has no logical dependencies with other restore
operations. This patch introduce some structures and some helper
functions for the target process to offload this task to the main CRIU
process.
Signed-off-by: Yanning Yang <yangyanning@sjtu.edu.cn>
When enabling parallel restore, the target process and the main CRIU
process need an IPC interface to communicate and transfer restore
commands. This patch adds a Unix domain TCP socket and stores this
socket in `fdstore`.
Signed-off-by: Yanning Yang <yangyanning@sjtu.edu.cn>
This functionality (#2527) is being reverted and excluded from this
release due to issue #2812.
It will be included in a subsequent release once all associated issues
are resolved.
Signed-off-by: Andrei Vagin <avagin@google.com>
This patch implements the entire logic to enable the offloading of
buffer object content restoration.
The goal of this patch is to offload the buffer object content
restoration to the main CRIU process so that this restoration can occur
in parallel with other restoration logic (mainly the restoration of
memory state in the restore blob, which is time-consuming) to speed up
the restore phase. The restoration of buffer object content usually
takes a significant amount of time for GPU applications, so
parallelizing it with other operations can reduce the overall restore
time.
It has three parts: the first replaces the restoration of buffer objects
in the target process by sending a parallel restore command to the main
CRIU process; the second implements the POST_FORKING hook in the amdgpu
plugin to enable buffer object content restoration in the main CRIU
process; the third stops the parallel thread in the RESUME_DEVICES_LATE
hook.
This optimization only focuses on the single-process situation (common
case). In other scenarios, it will turn to the original method. This is
achieved with the new `parallel_disabled` flag.
Signed-off-by: Yanning Yang <yangyanning@sjtu.edu.cn>
Currently the restore of buffer object comsumes a significant amount of
time. However, this part has no logical dependencies with other restore
operations. This patch introduce some structures and some helper
functions for the target process to offload this task to the main CRIU
process.
Signed-off-by: Yanning Yang <yangyanning@sjtu.edu.cn>
When enabling parallel restore, the target process and the main CRIU
process need an IPC interface to communicate and transfer restore
commands. This patch adds a Unix domain TCP socket and stores this
socket in `fdstore`.
Signed-off-by: Yanning Yang <yangyanning@sjtu.edu.cn>
The `goto interrupt` label is unnecessary as the code directly
returns after `cuda_process_checkpoint_action()`.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
When handing errors for functions such as `ptrace()`, `pipe()`, and
`fork()` it would be better to use `pr_perror` instead of `pr_err`
as it would include a message describing the encountered error.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
Temporarily disable CUDA plugin for `criu pre-dump`.
pre-dump currently fails with the following error:
Handling VMA with the following smaps entry: 1822c000-18da5000 rw-p 00000000 00:00 0 [heap]
Handling VMA with the following smaps entry: 200000000-200200000 ---p 00000000 00:00 0
Handling VMA with the following smaps entry: 200200000-200400000 rw-s 00000000 00:06 895 /dev/nvidia0
Error (criu/proc_parse.c:116): handle_device_vma plugin failed: No such file or directory
Error (criu/proc_parse.c:632): Can't handle non-regular mapping on 705693's map 200200000
Error (criu/cr-dump.c:1486): Collect mappings (pid: 705693) failed with -1
We plan to enable support for pre-dump by skipping nvidia mappings
in a separate patch.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
When creating a checkpoint of large models, the `checkpoint` action of
`cuda-checkpoint` can exceed the CRIU timeout. This causes CRIU to fail
with the following error, leaving the CUDA task in a locked state:
cuda_plugin: Checkpointing CUDA devices on pid 84145 restore_tid 84202
Error (criu/cr-dump.c:1791): Timeout reached. Try to interrupt: 0
Error (cuda_plugin.c:139): cuda_plugin: Unable to read output of cuda-checkpoint: Interrupted system call
Error (cuda_plugin.c:396): cuda_plugin: CHECKPOINT_DEVICES failed with
net: Unlock network
cuda_plugin: finished cuda_plugin stage 0 err -1
cuda_plugin: resuming devices on pid 84145
cuda_plugin: Restore thread pid 84202 found for real pid 84145
Unfreezing tasks into 1
Unseizing 84145 into 1
Error (criu/cr-dump.c:2111): Dumping FAILED.
To fix this, we set `task_info->checkpointed` before invoking
the `checkpoint` action to ensure that the CUDA task is resumed
even if CRIU times out.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
cuda-checkpoint returns the positive CUDA error code when it runs into an issue
and passing that along as the return value would cause errors to get ignored
Signed-off-by: Jesus Ramos <jeramos@nvidia.com>
If a CUDA process is already in a "locked" or "checkpointed" state
during criu dump, the CUDA plugin currently fails with an error because
it attempts an unnecessary "lock" action using the cuda-checkpoint tool.
This patch extends the CUDA plugin to handle such cases by first
verifying the initial state of the CUDA processes and skipping
unnecessary "lock" and "checkpoint" actions when a process has been
locked or checkpointed before CRIU is invoked.
In particular, CUDA tasks may already be in a "locked" or "checkpointed"
state to ensure consistent checkpoint/restore for distributed workloads,
such as model training, where multiple containers run across different
cluster nodes.
Another use case for this functionality is optimizing resource
utilization, where CUDA tasks with low-priority are preempted
immediately to release GPU resources needed by high-priority
tasks, and the paused workloads are later resumed or migrated
to another node.
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
The check for `/dev/nvidiactl` to determine if the CUDA plugin can be
used is unreliable because in some cases the default path for driver
installation is different [1]. This patch changes the logic to check
if a GPU device is available in `/proc/driver/nvidia/gpus/`. This
approach is similar to `torch.cuda.is_available()` and it is a more
accurate indicator.
The subsequent check for support of the `cuda-checkpoint --action`
option would confirm if the driver supports checkpoint/restore.
[1] https://github.com/NVIDIA/gpu-operatorFixes: #2509
Signed-off-by: Radostin Stoyanov <rstoyanov@fedoraproject.org>
