SWSPLAT-30708 xrt_kernel validate arg offset/size against execbuf payload

[stsoe]

Problem solved by the commit

hs_arg_setter::set_arg_value() and set_offset_value() write caller-supplied bytes into data + arg.offset() where data points into the 4096-byte exec BO payload. arg.offset() and arg.size() are parsed from the xclbin EMBEDDED_METADATA XML with no upper bound. A malicious xclbin with offset=0x10000 causes every set_arg() call to write past the mapped BO into adjacent heap memory (CWE-787, CVSS 7.1). fa_arg_setter has the same defect via fa_desc_offset().

Bug / issue (if any) fixed, which PR introduced the bug, how it was discovered

SWSPLAT-30708. Identified by Mythos AI security audit (MYTHOS-RyzenAI-XRT-H2). Parent finding: SWSPLAT-30705.

How problem was solved, alternative solutions (if any) and why they were rejected

Added payload_size() to run_impl that computes the true number of bytes available for kernel arguments in the command buffer:

reserved = bytes consumed by ert header + cu masks (computed from data ptr)
payload = kernel_command::allocation_size - reserved (physical ceiling)

if regmap_size > 0: cap to regmap_size bytes (validates it fits in physical)
if regmap_size == 0: unlimited args — use physical ceiling

This handles kernels where regmap_size is not set (unlimited arguments) while still bounding writes to the actual buffer.

The resulting payload_size is passed to each arg_setter at construction time as m_payload_size. Overflow-safe bounds checks are added before every copy_n in hs_arg_setter (set_arg_value, set_offset_value, get_arg_value) and before the fa_desc_offset pointer arithmetic in fa_arg_setter.

bo_cache.h is amended to expose bo_size as a public constexpr so kernel_command::allocation_size can reference it without a magic number.

All four make_arg_setter() sites (pl/ps/dpu kernels and mailbox_impl) updated.

Risks (if any) associated the changes in the commit

Low. The bounds check fires only when arg.offset() + count exceeds the payload — impossible for a well-formed xclbin. Malformed inputs that previously caused silent heap corruption now throw xrt_core::error(-EINVAL).

What has been tested and how, request additional testing if necessary

Code review against the SWSPLAT-30708 finding. Existing kernel argument tests exercise the nominal path. The AFL++ harness from the ticket (feed xclbin bytes to xrt::kernel + run.set_arg) is recommended to confirm the OOB write is closed.

[github-actions]

clang-tidy made some suggestions

[github-actions]

warning: class 'bo_cache_t' defines a non-default destructor but does not define a copy constructor, a copy assignment operator, a move constructor or a move assignment operator [cppcoreguidelines-special-member-functions]

class bo_cache_t
      ^