pkg/aflow/flow/patching: add bug fix patching workflow

2 files changed, 195 insertions, 0 deletions

diff --git a/pkg/aflow/flow/patching/base_commit.go b/pkg/aflow/flow/patching/base_commit.go
new file mode 100644
index 000000000..f49aba3d6
--- /dev/null
+++ b/pkg/aflow/flow/patching/base_commit.go
@@ -0,0 +1,57 @@
+// Copyright 2025 syzkaller project authors. All rights reserved.
+// Use of this source code is governed by Apache 2 LICENSE that can be found in the LICENSE file.
+
+package patching
+
+import (
+	"github.com/google/syzkaller/pkg/aflow"
+	"github.com/google/syzkaller/pkg/aflow/action/kernel"
+	"github.com/google/syzkaller/pkg/vcs"
+)
+
+var baseCommitPicker = aflow.NewFuncAction("base-commit-picker", pickBaseCommit)
+
+type baseCommitArgs struct {
+	// Can be used to override the selected base commit (for manual testing).
+	FixedBaseCommit string
+}
+
+type baseCommitResult struct {
+	KernelRepo   string
+	KernelCommit string
+}
+
+func pickBaseCommit(ctx *aflow.Context, args baseCommitArgs) (baseCommitResult, error) {
+	// Currently we use the latest RC of the mainline tree as the base.
+	// This is a reasonable choice overall in lots of cases, and it enables good caching
+	// of all artifacts (we need to rebuild them only approx every week).
+	// Potentially we can use subsystem trees for few important, well-maintained subsystems
+	// (mm, net, etc). However, it will work poorly for all subsystems. First, there is no
+	// machine-usable mapping of subsystems to repo/branch; second, lots of them are poorly
+	// maintained (can be much older than latest RC); third, it will make artifact caching
+	// much worse.
+	// In the future we ought to support automated rebasing of patches to requested trees/commits.
+	// We need it anyway, but it will also alleviate imperfect base commit picking.
+	const (
+		baseRepo   = "git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git"
+		baseBranch = "master"
+	)
+
+	res := baseCommitResult{
+		KernelRepo:   baseRepo,
+		KernelCommit: args.FixedBaseCommit,
+	}
+	if args.FixedBaseCommit != "" {
+		return res, nil
+	}
+
+	err := kernel.UseLinuxRepo(ctx, func(_ string, repo vcs.Repo) error {
+		head, err := repo.Poll(baseRepo, baseBranch)
+		if err != nil {
+			return err
+		}
+		res.KernelCommit, err = repo.ReleaseTag(head.Hash)
+		return err
+	})
+	return res, err
+}
diff --git a/pkg/aflow/flow/patching/patching.go b/pkg/aflow/flow/patching/patching.go
new file mode 100644
index 000000000..ef10f1a2e
--- /dev/null
+++ b/pkg/aflow/flow/patching/patching.go
@@ -0,0 +1,138 @@
+// Copyright 2025 syzkaller project authors. All rights reserved.
+// Use of this source code is governed by Apache 2 LICENSE that can be found in the LICENSE file.
+
+package patching
+
+import (
+	"encoding/json"
+
+	"github.com/google/syzkaller/pkg/aflow"
+	"github.com/google/syzkaller/pkg/aflow/action/crash"
+	"github.com/google/syzkaller/pkg/aflow/action/kernel"
+	"github.com/google/syzkaller/pkg/aflow/ai"
+	"github.com/google/syzkaller/pkg/aflow/tool/codesearcher"
+)
+
+type Inputs struct {
+	ReproOpts         string
+	ReproSyz          string
+	ReproC            string
+	KernelConfig      string
+	SyzkallerCommit   string
+	CodesearchToolBin string
+
+	// Same as in the manager config.
+	Syzkaller string
+	Image     string
+	Type      string
+	VM        json.RawMessage
+
+	// Use this fixed based kernel commit (for testing/local running).
+	FixedBaseCommit string
+}
+
+type Outputs struct {
+	PatchDescription string
+	PatchDiff        string
+}
+
+func init() {
+	tools := codesearcher.Tools
+
+	aflow.Register[Inputs, Outputs](
+		ai.WorkflowPatching,
+		"generate a kernel patch fixing a provided bug reproducer",
+		&aflow.Flow{
+			Root: &aflow.Pipeline{
+				Actions: []aflow.Action{
+					baseCommitPicker,
+					kernel.Checkout,
+					kernel.Build,
+					// Ensure we can reproduce the crash (and the build boots).
+					crash.Reproduce,
+					codesearcher.PrepareIndex,
+					&aflow.LLMAgent{
+						Name:        "debugger",
+						Reply:       "BugExplanation",
+						Temperature: 1,
+						Instruction: debuggingInstruction,
+						Prompt:      debuggingPrompt,
+						Tools:       tools,
+					},
+					&aflow.LLMAgent{
+						Name:        "diff-generator",
+						Reply:       "PatchDiff",
+						Temperature: 1,
+						Instruction: diffInstruction,
+						Prompt:      diffPrompt,
+						Tools:       tools,
+					},
+					&aflow.LLMAgent{
+						Name:        "description-generator",
+						Reply:       "PatchDescription",
+						Temperature: 1,
+						Instruction: descriptionInstruction,
+						Prompt:      descriptionPrompt,
+					},
+				},
+			},
+		},
+	)
+}
+
+// TODO: mention not doing assumptions about the source code, and instead querying code using tools.
+// TODO: mention to extensively use provided tools to confirm everything.
+// TODO: use cause bisection info, if available.
+
+const debuggingInstruction = `
+You are an experienced Linux kernel developer tasked with debugging a kernel crash root cause.
+You need to provide a detailed explanation of the root cause for another developer to be
+able to write a fix for the bug based on your explanation.
+Your final reply must contain only the explanation.
+
+Call some codesearch tools first.
+`
+
+const debuggingPrompt = `
+The crash is:
+
+{{.CrashReport}}
+`
+
+const diffInstruction = `
+You are an experienced Linux kernel developer tasked with creating a patch for a kernel bug.
+Your final reply should contain only the code diff in patch format.
+`
+
+const diffPrompt = `
+The crash that corresponds to the bug is:
+
+{{.CrashReport}}
+
+The explanation of the root cause of the bug is:
+
+{{.BugExplanation}}
+`
+
+const descriptionInstruction = `
+You are an experienced Linux kernel developer tasked with writing a commit description for
+a kernel bug fixing commit. The description should start with a one-line summary,
+and then include description of the bug being fixed, and how it's fixed by the provided patch.
+Your final reply should contain only the text of the commit description.
+Phrase the one-line summary so that it is not longer than 72 characters.
+The rest of the description must be word-wrapped at 72 characters.
+`
+
+const descriptionPrompt = `
+The crash that corresponds to the bug is:
+
+{{.CrashReport}}
+
+The explanation of the root cause of the bug is:
+
+{{.BugExplanation}}
+
+The diff of the bug fix is:
+
+{{.PatchDiff}}
+`