[libc] [compiler-rt] [clang-tools-extra] [llvm] [lldb] [flang] [clang] LLDB Debuginfod usage tests (with fixes) (PR #79181)

Kevin Frei via cfe-commits cfe-commits at lists.llvm.org
Fri Jan 26 09:46:45 PST 2024


================
@@ -0,0 +1,65 @@
+# Tests for basic Debuginfod functionality
+
+Because the Debuginfod protocol is a simple HTTP path-based system, one can
+mimic a Debuginfod server by setting up a directory structure to reflect the
+protocol properly. That's how all these tests operate. We override the default
+`DEBUGINFOD_URLS` property with a `file://` URL and populate it with the symbol
+files we need for testing.
+
+## What's being tested
+
+- For assumption validation, the `*-no-locator` tests verify that lldb works as
+  the test expects when files that Debuginfod should provide (`.dwp` files,
+  `.gnu.debuglink`'ed files, etc...) are _already_ there.
+- The `*-negative` tests validate that symbols _aren't_ found without
+  Debuginfod, to ensure they haven't been cached from previous runs (in the
+  hopes of preventing false positive testing).
+- The `*-positive*` tests check that the Debuginfod symbol locator is providing
+  the expected symbols when the debugger doesn't already have them available.
+
+### Symbol file variations tested
+
+There are 5 variations of symbol data where Debuginfod provides value:
+
+1. The `strip` build variation is a binary built with debug information (`-g`),
+   but stripped for deployment. The Debuginfod service can then host the
+   unstripped binary (as either `executable` or `debuginfo`).
+2. The `okdstrip` build variation is a binary build with `-g`, stripped for
+   deployment, where the Debuginfod service is hosting the output of
+   `objcopy --only-keep-debug` (which should also be linked to the stripped file
+   using `--add-gnu-debuglink`). Again, the file could be hosted as either
+   `executable` or `debuginfo`.
+3. The `split` build variation is a binary built with `-gsplit-dwarf` that
+   produces `.dwo` which are subsequently linked together (using `llvm-dwp`)
+   into a single `.dwp` file. The Debuginfod service hosts the `.dwp` file as
+   `debuginfo`.
+4. The `split-strip` build variation is a binary built with `-gsplit-dwarf`,
+   then stripped in the same manner as variation #1. For this variation,
+   Debuginfod hosts the unstripped binary as `executable` and the `.dwp` file as
+   `debuginfo`.
+5. The `split-okdstrip` build variation is the combination of variations 2 and
+   3, where Debuginfod hosts the `.gnu.debuglink`'ed file as `executable` and
+   the `.dwp` as `debuginfo`.
+
+### Lack of clarity/messy capabilities from Debuginfod
+
+The [debuginfod protocol](https://sourceware.org/elfutils/Debuginfod.html) is
+underspecified for some variations of symbol file deployment. The protocol
+itself is quite simple: query an HTTP server with the path
+`buildid/{.note.gnu.build-id hash}/debuginfo` or
+`buildid/{.note.gnu.build-id hash}/executable` to acquire "symbol data" or "the
+executable". Where there is lack of clarity, I prefer requesting `debuginfo`
+first, then falling back to `executable` (Scenarios #1 & #2). For Scenario #5,
+I've chosen to expect the stripped (i.e. not full) executable, which contains a
+number of sections necessary to correctly symbolicate will be hosted from the
+`executable` API. Depending upon how Debuginfod hosting services choose to
+support `.dwp` paired with stripped files, these assumptions may need to be
+revisited.
+
+I've also chosen to simply treat the `.dwp` file as `debuginfo` and the
+"only-keep-debug" stripped binary as `executable`. This scenario doesn't appear
+to work at all in GDB. Supporting it how I did seems more straight forward than
+trying to extend the protocol. The protocol _does_ support querying for section
+contents by name for a given build ID, but adding support for that in LLDB
+looks...well beyond my current capability (and LLVM's Debuginfod library doesn't
+support it at this writing, anyway).
----------------
kevinfrei wrote:

Changed voicing

https://github.com/llvm/llvm-project/pull/79181


More information about the cfe-commits mailing list