Fix crash in xdna_aie_array initialization

[bisingha-xilinx]

Problem and fix for https://jira.xilinx.com/browse/AIESW-21965:

While running multi process stress tests severe crash occurred. Earlier while initializing xdna_aie_array, a debug API XAie_GetPartitionFdList() was being used and ve2 userspace used to leak FDs and resources were not getting freed properly which led to crash in stress tests.

Rootcauses:

1. FD leaks in xdna_aie_array initialization.
2. invalid num_cols while getting partition info while running multi process tests.
3. xdna_aie_array destruction used to happen after the hwctx destruction which led to improper destruction of hwctx.

Fixes:

1. Fixed the FD leaks by adding a new param in DRM_IOCTL_AMDXDNA_GET_ARRAY to get the correct and single FD for a specific aie partition instead of getting all the FDs for all the partitions.
2. Added a check to get the correct num_cols for a correct hwctx for a specific PID to avoid getting wrong num_col in multi process tests.
3. Destroying xdna_aie_array explicity before destroying the hw context.

Tested:

1. The provided test from https://jira.xilinx.com/browse/AIESW-21965. Tested for 1000 iterations by having 7 instances of the host running parallelly in the background. With the above fix not seeing any crash anymore. And also getting correct refcounts for zocl, amdxdna, xilinx_aie in "lsmod". And able to load-unload the modules.
2. Tested all the basic XRT hw tests in VE2.

[Copilot]

Pull request overview

This PR fixes a critical crash in xdna_aie_array initialization that occurred during multi-process stress tests. The crash was caused by file descriptor leaks from using a debug API, incorrect partition column counts in multi-process scenarios, and improper cleanup order during context destruction.

Changes:

• Replaced debug API XAie_GetPartitionFdList() with a new kernel IOCTL DRM_AMDXDNA_HWCTX_AIE_PART_FD to obtain a single correct FD without leaks
• Added PID-based filtering when retrieving partition info to ensure correct column counts in multi-process tests
• Moved xdna_aie_array destruction to occur before hardware context destruction for proper cleanup order

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
src/shim_ve2/xdna_hwctx.cpp	Added PID check for partition info retrieval and explicit aie_array destruction before hwctx cleanup
src/shim_ve2/xdna_aie_array.h	Added private method get_aie_partition_fd for FD retrieval
src/shim_ve2/xdna_aie_array.cpp	Replaced debug API with new helper method using kernel IOCTL to get partition FD
src/include/uapi/drm_local/amdxdna_accel.h	Added new parameter DRM_AMDXDNA_HWCTX_AIE_PART_FD for IOCTL interface
src/driver/amdxdna/ve2_debug.c	Implemented kernel-side handler for new AIE partition FD retrieval IOCTL

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The function uses args->num_element to pass the hwctx handle, which is semantically unclear. Consider using a more descriptive field name or adding a comment explaining this non-obvious usage of num_element as a handle parameter.

[bisingha-xilinx]

Making use of the exiting framework.

[Copilot]

The file descriptor is closed on copy_to_user failure, but if copy_to_user succeeds, the FD is transferred to userspace without being tracked. This could lead to FD leaks if the userspace process terminates unexpectedly before using or closing the FD. Consider using get_unused_fd_flags with fd_install pattern for proper FD lifecycle management.

[bisingha-xilinx]

FD is getting closed from the XAie_Finish() which is being called in ~xdna_aie_array()

[maxzhen]

What is partition FD? Is it an ID for a single partition? Why would application care about this ID? How does it map to aie2/aie4 platforms?

[chvamshi-xilinx]

What is partition FD? Is it an ID for a single partition? Why would application care about this ID? How does it map to aie2/aie4 platforms?

Yes Max. This FD represents single partition. This is required for XDP (Profiling use case) in AIE2PS. This is not required in aie2/aie4 platforms.

[maxzhen]

What is partition FD? Is it an ID for a single partition? Why would application care about this ID? How does it map to aie2/aie4 platforms?

Yes Max. This FD represents single partition. This is required for XDP (Profiling use case) in AIE2PS. This is not required in aie2/aie4 platforms.

Why we can't use ctx ID for this? Each ctx can only be created on one partition for its entire life cycle, right? So, one ctx maps to one partition, no?

[chvamshi-xilinx]

What is partition FD? Is it an ID for a single partition? Why would application care about this ID? How does it map to aie2/aie4 platforms?

Yes Max. This FD represents single partition. This is required for XDP (Profiling use case) in AIE2PS. This is not required in aie2/aie4 platforms.

Why we can't use ctx ID for this? Each ctx can only be created on one partition for its entire life cycle, right? So, one ctx maps to one partition, no?

This FD is created by AIE driver, We get this FD by calling aie_partition_get_fd function
https://github.com/amd/xdna-driver/pull/990/changes#diff-3eb91d41ff4f6a0011e67d0f849aede12e1abd5726cc7aa0f103b082d740e91bR706
Initially we were calling AIE driver userspace API, that was causing some issues. This is the standard way to get aie partition fd from AIE driver.