Conversation
The existing PGO cycle bundles config-only, syscall-prune, and layout-ordering candidates into a single rebuild loop. When all we want is the syscall-prune verdict (e.g. tightening the prune table after a userspace change), paying for the layout-ordering detour is wasteful, and the QEMU trace + analyze step gets re-run on every invocation. build.sh grows kernel_syscall_prune_cycle alongside the existing kernel_pgo_cycle: same primitives (build_candidate_kernel, boot_not_regressed, restore_kernel_artifacts), but only baseline / config-only / syscall-prune candidates, defaulting QEMU_LOG to exec,cpu,in_asm so the R7 dumps needed for syscall extraction are present. The stage selects the smallest linux.axf that does not regress shell_ready_ms, mirroring the existing cycle's selection rule. prepare_kernel_profile_analysis extracts the trace-collect+analyze pair into a shared helper keyed on (IMAGE_FP, sha256(workload), trace selector). Both cycles share the namespace but resolve to distinct cache directories via the trace tag, so PGO's exec,in_asm trace and the new cycle's exec,cpu,in_asm trace do not collide. IMAGE_FP already covers scripts/, configs/, patches/, and build.sh, so any change that affects the baseline binary or the analysis tooling invalidates the cache. materialize_cache_tree uses cp -f rather than hardlinks: link_cached_tree exposes cache files via symlinks in working dirs, and a hardlink would let any future writer that follows that symlink mutate the cache through a shared inode. cp -f makes the cache the canonical copy. scripts/qemu-trace-to-orderfile.py rewrites the syscall pairing state machine to match QEMU's actual -d cpu ordering: register dumps precede the TB whose entry state they capture. The previous code kept a backward-binding fallback that, when a TB without a preceding regdump showed up (orphan TB), bound the next R07 to the orphan instead of holding it for the upcoming TB. In a real trace this silently misattributes or drops syscalls and feeds a wrong prune table. pending_tb is dropped entirely; an R07 always binds forward to the next TB or stays pending until one arrives. scripts/test_qemu_trace_to_orderfile.py covers the happy path, the orphan-TB regression (which fails under the old backward-binding), and the MAX_ARM_SYSCALL boundary at 511/512. Three Python script invocations gain an explicit python3 prefix. generate-syscall-prune-table.py is committed as 100644, so calling it via the bare path raised Permission denied; the other two are defensive consistency.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The existing PGO cycle bundles config-only, syscall-prune, and layout-ordering candidates into a single rebuild loop. When all we want is the syscall-prune verdict (e.g. tightening the prune table after a userspace change), paying for the layout-ordering detour is wasteful, and the QEMU trace + analyze step gets re-run on every invocation.
build.sh grows kernel_syscall_prune_cycle alongside the existing kernel_pgo_cycle: same primitives (build_candidate_kernel, boot_not_regressed, restore_kernel_artifacts), but only baseline / config-only / syscall-prune candidates, defaulting QEMU_LOG to exec,cpu,in_asm so the R7 dumps needed for syscall extraction are present. The stage selects the smallest linux.axf that does not regress shell_ready_ms, mirroring the existing cycle's selection rule.
prepare_kernel_profile_analysis extracts the trace-collect+analyze pair into a shared helper keyed on (IMAGE_FP, sha256(workload), trace selector). Both cycles share the namespace but resolve to distinct cache directories via the trace tag, so PGO's exec,in_asm trace and the new cycle's exec,cpu,in_asm trace do not collide. IMAGE_FP already covers scripts/, configs/, patches/, and build.sh, so any change that affects the baseline binary or the analysis tooling invalidates the cache.
materialize_cache_tree uses cp -f rather than hardlinks: link_cached_tree exposes cache files via symlinks in working dirs, and a hardlink would let any future writer that follows that symlink mutate the cache through a shared inode. cp -f makes the cache the canonical copy.
scripts/qemu-trace-to-orderfile.py rewrites the syscall pairing state machine to match QEMU's actual -d cpu ordering: register dumps precede the TB whose entry state they capture. The previous code kept a backward-binding fallback that, when a TB without a preceding regdump showed up (orphan TB), bound the next R07 to the orphan instead of holding it for the upcoming TB. In a real trace this silently misattributes or drops syscalls and feeds a wrong prune table. pending_tb is dropped entirely; an R07 always binds forward to the next TB or stays pending until one arrives.
scripts/test_qemu_trace_to_orderfile.py covers the happy path, the orphan-TB regression (which fails under the old backward-binding), and the MAX_ARM_SYSCALL boundary at 511/512.
Three Python script invocations gain an explicit python3 prefix. generate-syscall-prune-table.py is committed as 100644, so calling it via the bare path raised Permission denied; the other two are defensive consistency.