Re: [PATCH] kallsyms: strip LTO suffixes from static functions

From: Nick Desaulniers
Date: Mon Jun 28 2021 - 13:54:44 EST


On Tue, Jun 22, 2021 at 1:18 PM Fangrui Song <maskray@xxxxxxxxxx> wrote:
>
> On 2021-06-22, 'Nick Desaulniers' via Clang Built Linux wrote:
> >Similar to:
> >commit 8b8e6b5d3b01 ("kallsyms: strip ThinLTO hashes from static
> >functions")
> >
> >It's very common for compilers to modify the symbol name for static
> >functions as part of optimizing transformations. That makes hooking
> >static functions (that weren't inlined or DCE'd) with kprobes difficult.
> >
> >Full LTO uses a different mangling scheme than thin LTO; full LTO
> >imports all code into effectively one big translation unit. It must
> >rename static functions to prevent collisions. Strip off these suffixes
> >so that we can continue to hook such static functions.
>
> See below. The message needs a change.
>
> I can comment on the LTO side thing, but a maintainer needs to check
> about the kernel side logic.
>
> Reviewed-by: Fangrui Song <maskray@xxxxxxxxxx>
>
> >Reported-by: KE.LI(Lieke) <like1@xxxxxxxx>
> >Tested-by: KE.LI(Lieke) <like1@xxxxxxxx>
> >Signed-off-by: Nick Desaulniers <ndesaulniers@xxxxxxxxxx>
> >---
> > kernel/kallsyms.c | 18 ++++++++++++++++++
> > 1 file changed, 18 insertions(+)
> >
> >diff --git a/kernel/kallsyms.c b/kernel/kallsyms.c
> >index 4067564ec59f..14cf3a6474de 100644
> >--- a/kernel/kallsyms.c
> >+++ b/kernel/kallsyms.c
> >@@ -188,6 +188,24 @@ static inline bool cleanup_symbol_name(char *s)
> >
> > return res != NULL;
> > }
> >+#elif defined(CONFIG_LTO_CLANG_FULL)
> >+/*
> >+ * LLVM mangles static functions for full LTO so that two static functions with
> >+ * the same identifier do not collide when all code is combined into one
> >+ * module. The scheme used converts references to foo into
> >+ * foo.llvm.974640843467629774, for example. This can break hooking of static
> >+ * functions with kprobes.
> >+ */
>
> The comment should say ThinLTO instead.
>
> The .llvm.123 suffix is for global scope promotion for local linkage
> symbols. The scheme is ThinLTO specific. This ensures that a local

Oh, boy. Indeed. I had identified the mangling coming from
getGlobalNameForLocal(), but looking at the call chain now I see:

FunctionImportGlobalProcessing::processGlobalForThinLTO()
-> FunctionImportGlobalProcessing::getPromotedName()
-> ModuleSummaryIndex::getGlobalNameForLocal()

I'm not sure then how I figured it was specific to full LTO.

Android recently switched from thin LTO to full LTO, which is what I
assumed was the cause of the bug report. Rereading our internal bug
report, it was tested against a prior version that did the symbol
truncation for thinLTO. I then assumed this was full LTO specific for
whatever reason, and modified the patch to only apply to full LTO. I
see via the above call chain that this patch is not correct. Let me
send my original patch as a v2. b/189560201 if you're interested.

> linkage symbol, when imported into multiple translation units, then
> compiled into different object files, during linking, the copies can be
> deduplicated. This matters for code size and for correctness when the
> function address is taken.
>
> Regular LTO (sometimes called full LTO) uses the regular name.\d+
> scheme.
>
> >+static inline bool cleanup_symbol_name(char *s)
> >+{
> >+ char *res;
> >+
> >+ res = strstr(s, ".llvm.");
> >+ if (res)
> >+ *res = '\0';
> >+
> >+ return res != NULL;
> >+}
> > #else
> > static inline bool cleanup_symbol_name(char *s) { return false; }
> > #endif
> >--
> >2.32.0.288.g62a8d224e6-goog
>
> I wonder whether it makes sense to strip all `.something` suffixes.
> For example, the recent -funique-internal-linkage-name (which can
> improve sample profile accuracy) uses the `.__uniq.1234` scheme.
>
> Function specialization/clones can create arbitrary `.123` suffixes.

I definitely don't see hooking static functions via kprobes as being
scalable. There are numerous different mangling schemes different
compilers apply to different static functions.

--
Thanks,
~Nick Desaulniers