Re: [PATCH] compiler: Simplify generic RELOC_HIDE()

[Peter Zijlstra]

On Mon, Mar 23, 2026 at 04:25:52PM -0700, Nathan Chancellor wrote:
> On Thu, Mar 19, 2026 at 02:52:38PM +0100, Marco Elver wrote:
> > When enabling Context Analysis (CONTEXT_ANALYSIS := y) in arch/x86/kvm
> > code, Clang's Thread Safety Analysis failed to recognize that identical
> > per_cpu() accesses refer to the same lock:
> > 
> > |   CC [M]  arch/x86/kvm/vmx/posted_intr.o
> > | arch/x86/kvm/vmx/posted_intr.c:186:2: error: releasing raw_spinlock '__ptr + __per_cpu_offset[vcpu->cpu]' that was not held [-Werror,-Wthread-safety-analysis]
> > |   186 |         raw_spin_unlock(&per_cpu(wakeup_vcpus_on_cpu_lock, vcpu->cpu));
> > |       |         ^
> > | ./include/linux/spinlock.h:276:32: note: expanded from macro 'raw_spin_unlock'
> > |   276 | #define raw_spin_unlock(lock)           _raw_spin_unlock(lock)
> > |       |                                         ^
> > | arch/x86/kvm/vmx/posted_intr.c:207:1: error: raw_spinlock '__ptr + __per_cpu_offset[vcpu->cpu]' is still held at the end of function [-Werror,-Wthread-safety-analysis]
> > |   207 | }
> > |       | ^
> > | arch/x86/kvm/vmx/posted_intr.c:182:2: note: raw_spinlock acquired here
> > |   182 |         raw_spin_lock_nested(&per_cpu(wakeup_vcpus_on_cpu_lock, vcpu->cpu),
> > |       |         ^
> > | ./include/linux/spinlock.h:235:2: note: expanded from macro 'raw_spin_lock_nested'
> > |   235 |         _raw_spin_lock(((void)(subclass), (lock)))
> > |       |         ^
> > | 2 errors generated.
> > 
> > This occurred because the default RELOC_HIDE() implementation (used by
> > the per-CPU macros) is a statement expression containing an intermediate
> > 'unsigned long' variable (this version appears to predate Git history).
> > 
> > While the analysis strips away inner casts when resolving pointer
> > aliases, it stops when encountering intermediate non-pointer variables
> > (this is Thread Safety Analysis specific and irrelevant for codegen).
> > This prevents the analysis from concluding that the pointers passed to
> > e.g. raw_spin_lock() and raw_spin_unlock() were identical when per-CPU
> > accessors are used.
> > 
> > Simplify RELOC_HIDE() to a single expression. This preserves the intent
> > of obfuscating UB-introducing out-of-bounds pointer calculations from
> > the compiler via the 'unsigned long' cast, but allows the alias analysis
> > to successfully resolve the pointers.
> > 
> > Using a recent Clang version, I observe that generated code remains the
> > same for vmlinux; the intermediate variable was already being optimized
> > away (for any respectable modern compiler, not doing so would be an
> > optimizer bug). Note that GCC provides its own version of RELOC_HIDE(),
> > so this change only affects Clang builds.
> > 
> > Add a test case to lib/test_context-analysis.c to catch any regressions.
> > 
> > Link: https://lore.kernel.org/all/e3946223-4543-4a76-a328-9c6865e95192@xxxxxxx/
> > Reported-by: Bart Van Assche <bvanassche@xxxxxxx>
> > Reported-by: Sean Christopherson <seanjc@xxxxxxxxxx>
> > Signed-off-by: Marco Elver <elver@xxxxxxxxxx>
> 
> Reviewed-by: Nathan Chancellor <nathan@xxxxxxxxxx>

Thanks, I'll go stick it in tip/locking/core.