| 1 | // SPDX-License-Identifier: AGPL-3.0-or-later |
| 2 | |
| 3 | package markdown |
| 4 | |
| 5 | import ( |
| 6 | "regexp" |
| 7 | "sync" |
| 8 | |
| 9 | "github.com/microcosm-cc/bluemonday" |
| 10 | ) |
| 11 | |
| 12 | // sanitizer is the package-private bluemonday policy. Built once; |
| 13 | // safe for concurrent use. |
| 14 | // |
| 15 | // Threat model: |
| 16 | // |
| 17 | // - Goldmark with HTML rendering off escapes user-injected raw |
| 18 | // HTML at parse time, so the sanitizer is *defense in depth* — |
| 19 | // anything that reaches it should already be Goldmark-emitted |
| 20 | // HTML or our own AST-extension output. |
| 21 | // - The strict scheme allowlist (`http`, `https`, `mailto`) blocks |
| 22 | // `javascript:`, `data:`, `vbscript:` URIs entirely. |
| 23 | // - We diverge from GitHub by *not* allowing `data:image/...`. If |
| 24 | // a user wants an inline image, repo-relative paths work via |
| 25 | // /raw/. Documented in docs/markdown.md. |
| 26 | var sanitizer = func() *bluemonday.Policy { |
| 27 | p := bluemonday.UGCPolicy() |
| 28 | |
| 29 | // Headings keep their auto-generated id so anchor links work. |
| 30 | p.AllowAttrs("id").OnElements("h1", "h2", "h3", "h4", "h5", "h6") |
| 31 | |
| 32 | // Code-block class allowlist for Chroma (`language-foo`). The |
| 33 | // SpaceSeparatedTokens matcher is bluemonday-built-in; we |
| 34 | // further constrain via the regex below so only `language-*` and |
| 35 | // chroma's own ancillary classes pass. |
| 36 | p.AllowAttrs("class").Matching(reCodeClass).OnElements("code", "pre", "span") |
| 37 | |
| 38 | // GFM task lists: Goldmark emits `<input checked="" disabled="" |
| 39 | // type="checkbox" />` inside `<li>`. UGCPolicy doesn't allow |
| 40 | // input by default; whitelist the disabled-checkbox shape only. |
| 41 | // `type` is matched against "checkbox"; `disabled` and `checked` |
| 42 | // are HTML boolean attrs (value commonly empty), so we don't |
| 43 | // constrain the value — presence is the signal. |
| 44 | p.AllowAttrs("type").Matching(regexp.MustCompile(`^checkbox$`)).OnElements("input") |
| 45 | p.AllowAttrs("disabled", "checked").OnElements("input") |
| 46 | |
| 47 | // Folded sections + keyboard markers. Common in READMEs. |
| 48 | p.AllowElements("details", "summary", "kbd", "sup", "sub") |
| 49 | |
| 50 | // Mention / ref / commit anchors emitted by our extensions carry |
| 51 | // these classes for styling. The base UGC policy already allows |
| 52 | // <a> with href + rel; we just need the class allowlist above to |
| 53 | // cover `shithub-mention`, `shithub-ref`, `shithub-commit`. |
| 54 | |
| 55 | // Hard-restrict URL schemes. UGCPolicy already restricts schemes |
| 56 | // on <a>, but we tighten further: drop ftp, drop data:, leave |
| 57 | // only http(s) + mailto. |
| 58 | p.AllowURLSchemes("http", "https", "mailto") |
| 59 | // Image schemes — UGC allows http/https only by default; we keep |
| 60 | // that. No data: anywhere. |
| 61 | p.AllowImages() |
| 62 | |
| 63 | // rel="noopener noreferrer" auto-added when target="_blank" is |
| 64 | // set. We only set target via opts on autolinks; let bluemonday |
| 65 | // keep its default rel-handling. |
| 66 | p.RequireNoFollowOnLinks(true) |
| 67 | return p |
| 68 | }() |
| 69 | |
| 70 | var reCodeClass = regexp.MustCompile(`^(?:language-[A-Za-z0-9_+\-]+|chroma|chroma-[a-zA-Z]+|nl|ln|line|hl)(?:\s+(?:language-[A-Za-z0-9_+\-]+|chroma|chroma-[a-zA-Z]+|nl|ln|line|hl))*$`) |
| 71 | |
| 72 | // sanitizeBytes is the hot-path entry the Render pipeline uses. The |
| 73 | // bluemonday Policy is itself goroutine-safe; the once.Do here keeps |
| 74 | // the regex compilation cost off the first request path. |
| 75 | var sanitizerOnce sync.Once |
| 76 | |
| 77 | func sanitizeBytes(in []byte) []byte { |
| 78 | sanitizerOnce.Do(func() { |
| 79 | // Touch the package-level sanitizer to ensure it's built. |
| 80 | _ = sanitizer |
| 81 | }) |
| 82 | return sanitizer.SanitizeBytes(in) |
| 83 | } |
| 84 |