pcie_capability_read_dword leaves *dst uninitialized on register-not-implemented path diverging from LinuxKPI contract

Summary

pcie_capability_read_dword() returns -EINVAL without writing *dst when pcie_capability_reg_implemented() is false (linux_pci.c:35-36). Upstream Linux instead does *val=0 return 0. LinuxKPI consumers ported from upstream follow upstream convention ignore return value then read *dst unconditionally consuming uninitialized kernel-stack memory. Two live amdgpu call sites affected: pcie_get_width_cap() (pci.h:612) feeding amdgpu_device.c:3550 sets adev->pm.pcie_mlw_mask from uninitialized-derived link_width and amdgpu/cik.c:1634 gates clk_req_support on stale stack bit. Trigger is genuine runtime predicate: pcie_capability_reg_implemented returns false for PCI_EXP_LNKCAP whenever pcie_cap_has_lnkctl() false i.e. for PCIe v1 device whose type is not ROOT_PORT/ENDPOINT/LEG_END (switch upstream/downstream ports PCIe bridges). GPU upstream bridge can be such device on PCIe 1.x systems with discrete GPU behind switch. Impact: incorrect driver behavior potential hardware mis-programming possible local DoS from bad register write. No proven path to arbitrary code execution or direct userspace info leak uninitialized word masked to few bits before use.